diff --git a/.pylintrc b/.pylintrc
index 37b812ffe..f541638e7 100644
--- a/.pylintrc
+++ b/.pylintrc
@@ -31,6 +31,7 @@ ignore-paths=/usr/lib/.*$,
              modules/xadapter,
              modules/meissonic,
              modules/omnigen,
+             modules/instantir,
              modules/pulid/eva_clip,
              repositories,
              extensions-builtin/sd-webui-agent-scheduler,
diff --git a/.ruff.toml b/.ruff.toml
index 28439c73c..154d9a2c1 100644
--- a/.ruff.toml
+++ b/.ruff.toml
@@ -26,6 +26,7 @@ exclude = [
     "modules/xadapter",
     "modules/meissonic",
     "modules/omnigen",
+    "modules/instantir",
     "modules/pulid/eva_clip",
     "repositories",
     "extensions-builtin/sd-extension-chainner/nodes",
diff --git a/CHANGELOG.md b/CHANGELOG.md
index ee7199296..4bb10f5de 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -15,6 +15,11 @@ This release can be considered an LTS release before we kick off the next round
   - select in *scripts -> pulid*  
   - compatible with *sdxl*  
   - can be used in xyz grid  
+- [InstantIR](https://github.com/instantX-research/InstantIR): Blind Image Restoration with Instant Generative Reference
+  - alternative to traditional `img2img` with more control over restoration process  
+  - select in *image -> scripts -> instantir*  
+  - compatible with *sdxl*  
+  - *note*: after used once it cannot be unloaded without reloading base model  
 - SD3: ControlNets:  
   - *InstantX Canny, Pose, Depth, Tile*  
   - *Alimama Inpainting, SoftEdge*  
@@ -40,13 +45,12 @@ This release can be considered an LTS release before we kick off the next round
 - CLI:  
   - refactor command line params  
     run `webui.sh`/`webui.bat` with `--help` to see all options  
+  - added `cli/model-metadata.py` to display metadata in any safetensors file
+  - added `cli/model-keys.py` to quicky display content of any safetensors file
 - Other:  
   - Model loader: Report modules included in safetensors when attempting to load a model  
   - Repo: move screenshots to GH pages  
   - Requirements: update  
-- CLI:
-  - added `cli/model-metadata.py` to display metadata in any safetensors file
-  - added `cli/model-keys.py` to quicky display content of any safetensors file
 - Fixes:  
   - custom watermark add alphablending  
   - detailer min/max size as fractions of image size  
diff --git a/modules/img2img.py b/modules/img2img.py
index faf65161a..f3bec5b02 100644
--- a/modules/img2img.py
+++ b/modules/img2img.py
@@ -267,6 +267,7 @@ def img2img(id_task: str, state: str, mode: int,
         processed = modules.scripts.scripts_img2img.run(p, *args)
         if processed is None:
             processed = processing.process_images(p)
+        processed = modules.scripts.scripts_img2img.after(p, processed, *args)
     p.close()
     generation_info_js = processed.js() if processed is not None else ''
     if processed is None:
diff --git a/modules/instantir/__init__.py b/modules/instantir/__init__.py
new file mode 100644
index 000000000..fdd6cf8a0
--- /dev/null
+++ b/modules/instantir/__init__.py
@@ -0,0 +1,3 @@
+from .sdxl_instantir import InstantIRPipeline
+from .lcm_single_step_scheduler import LCMSingleStepScheduler
+from .ip_adapter.utils import init_adapter_in_unet, load_adapter_to_pipe
diff --git a/modules/instantir/aggregator.py b/modules/instantir/aggregator.py
new file mode 100644
index 000000000..fd6151003
--- /dev/null
+++ b/modules/instantir/aggregator.py
@@ -0,0 +1,983 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders.single_file_model import FromOriginalModelMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import TextImageProjection, TextImageTimeEmbedding, TextTimeEmbedding, TimestepEmbedding, Timesteps
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.unets.unet_2d_blocks import (
+    CrossAttnDownBlock2D,
+    DownBlock2D,
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+)
+from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class ZeroConv(nn.Module):
+    def __init__(self, label_nc, norm_nc, mask=False):
+        super().__init__()
+        self.zero_conv = zero_module(nn.Conv2d(label_nc+norm_nc, norm_nc, 1, 1, 0))
+        self.mask = mask
+
+    def forward(self, hidden_states, h_ori=None):
+        # with torch.cuda.amp.autocast(enabled=False, dtype=torch.float32):
+        c, h = hidden_states
+        if not self.mask:
+            h = self.zero_conv(torch.cat([c, h], dim=1))
+        else:
+            h = self.zero_conv(torch.cat([c, h], dim=1)) * torch.zeros_like(h)
+        if h_ori is not None:
+            h = torch.cat([h_ori, h], dim=1)
+        return h
+
+
+class SFT(nn.Module):
+    def __init__(self, label_nc, norm_nc, mask=False):
+        super().__init__()
+
+        # param_free_norm_type = str(parsed.group(1))
+        ks = 3
+        pw = ks // 2
+
+        self.mask = mask
+
+        nhidden = 128
+
+        self.mlp_shared = nn.Sequential(
+            nn.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw),
+            nn.SiLU()
+        )
+        self.mul = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw)
+        self.add = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw)
+
+    def forward(self, hidden_states, mask=False):
+        
+        c, h = hidden_states
+        mask = mask or self.mask
+        assert mask is False
+
+        actv = self.mlp_shared(c)
+        gamma = self.mul(actv)
+        beta = self.add(actv)
+
+        if self.mask:
+            gamma = gamma * torch.zeros_like(gamma)
+            beta = beta * torch.zeros_like(beta)
+        # gamma_ori, gamma_res = torch.split(gamma, [h_ori_c, h_c], dim=1)
+        # beta_ori, beta_res = torch.split(beta, [h_ori_c, h_c], dim=1)
+        # print(gamma_ori.mean(), gamma_res.mean(), beta_ori.mean(), beta_res.mean())
+        h = h * (gamma + 1) + beta
+        # sample_ori, sample_res = torch.split(h, [h_ori_c, h_c], dim=1)
+        # print(sample_ori.mean(), sample_res.mean())
+
+        return h
+
+
+@dataclass
+class AggregatorOutput(BaseOutput):
+    """
+    The output of [`Aggregator`].
+
+    Args:
+        down_block_res_samples (`tuple[torch.Tensor]`):
+            A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
+            be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
+            used to condition the original UNet's downsampling activations.
+        mid_down_block_re_sample (`torch.Tensor`):
+            The activation of the midde block (the lowest sample resolution). Each tensor should be of shape
+            `(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
+            Output can be used to condition the original UNet's middle block activation.
+    """
+
+    down_block_res_samples: Tuple[torch.Tensor]
+    mid_block_res_sample: torch.Tensor
+
+
+class ConditioningEmbedding(nn.Module):
+    """
+    Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
+    [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
+    training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
+    convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
+    (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
+    model) to encode image-space conditions ... into feature maps ..."
+    """
+
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        conditioning_channels: int = 3,
+        block_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
+    ):
+        super().__init__()
+
+        self.conv_in = nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
+
+        self.blocks = nn.ModuleList([])
+
+        for i in range(len(block_out_channels) - 1):
+            channel_in = block_out_channels[i]
+            channel_out = block_out_channels[i + 1]
+            self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
+            self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
+
+        self.conv_out = zero_module(
+            nn.Conv2d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
+        )
+
+    def forward(self, conditioning):
+        embedding = self.conv_in(conditioning)
+        embedding = F.silu(embedding)
+
+        for block in self.blocks:
+            embedding = block(embedding)
+            embedding = F.silu(embedding)
+
+        embedding = self.conv_out(embedding)
+
+        return embedding
+
+
+class Aggregator(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+    """
+    Aggregator model.
+
+    Args:
+        in_channels (`int`, defaults to 4):
+            The number of channels in the input sample.
+        flip_sin_to_cos (`bool`, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, defaults to 0):
+            The frequency shift to apply to the time embedding.
+        down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
+        block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, defaults to 2):
+            The number of layers per block.
+        downsample_padding (`int`, defaults to 1):
+            The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, defaults to 1):
+            The scale factor to use for the mid block.
+        act_fn (`str`, defaults to "silu"):
+            The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32):
+            The number of groups to use for the normalization. If None, normalization and activation layers is skipped
+            in post-processing.
+        norm_eps (`float`, defaults to 1e-5):
+            The epsilon to use for the normalization.
+        cross_attention_dim (`int`, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
+            The dimension of the attention heads.
+        use_linear_projection (`bool`, defaults to `False`):
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        num_class_embeds (`int`, *optional*, defaults to 0):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        upcast_attention (`bool`, defaults to `False`):
+        resnet_time_scale_shift (`str`, defaults to `"default"`):
+            Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
+        projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
+            The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
+            `class_embed_type="projection"`.
+        controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+            The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+        conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
+            The tuple of output channel for each block in the `conditioning_embedding` layer.
+        global_pool_conditions (`bool`, defaults to `False`):
+            TODO(Patrick) - unused parameter.
+        addition_embed_type_num_heads (`int`, defaults to 64):
+            The number of heads to use for the `TextTimeEmbedding` layer.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 4,
+        conditioning_channels: int = 3,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str, ...] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+        layers_per_block: int = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: int = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int, ...]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+        global_pool_conditions: bool = False,
+        addition_embed_type_num_heads: int = 64,
+        pad_concat: bool = False,
+    ):
+        super().__init__()
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+        self.pad_concat = pad_concat
+
+        # Check inputs
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        # input
+        conv_in_kernel = 3
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        # time
+        time_embed_dim = block_out_channels[0] * 4
+        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+        timestep_input_dim = block_out_channels[0]
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+        )
+
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
+        elif addition_embed_type is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+        # control net conditioning embedding
+        self.ref_conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        self.down_blocks = nn.ModuleList([])
+        self.controlnet_down_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+
+        # controlnet_block = ZeroConv(output_channel, output_channel)
+        controlnet_block = nn.Sequential(
+            SFT(output_channel, output_channel),
+            zero_module(nn.Conv2d(output_channel, output_channel, kernel_size=1))
+        )
+        self.controlnet_down_blocks.append(controlnet_block)
+
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block,
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim,
+                num_attention_heads=num_attention_heads[i],
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                downsample_padding=downsample_padding,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+            )
+            self.down_blocks.append(down_block)
+
+            for _ in range(layers_per_block):
+                # controlnet_block = ZeroConv(output_channel, output_channel)
+                controlnet_block = nn.Sequential(
+                    SFT(output_channel, output_channel),
+                    zero_module(nn.Conv2d(output_channel, output_channel, kernel_size=1))
+                )
+                self.controlnet_down_blocks.append(controlnet_block)
+
+            if not is_final_block:
+                # controlnet_block = ZeroConv(output_channel, output_channel)
+                controlnet_block = nn.Sequential(
+                    SFT(output_channel, output_channel),
+                    zero_module(nn.Conv2d(output_channel, output_channel, kernel_size=1))
+                )
+                self.controlnet_down_blocks.append(controlnet_block)
+
+        # mid
+        mid_block_channel = block_out_channels[-1]
+
+        # controlnet_block = ZeroConv(mid_block_channel, mid_block_channel)
+        controlnet_block = nn.Sequential(
+            SFT(mid_block_channel, mid_block_channel),
+            zero_module(nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1))
+        )
+        self.controlnet_mid_block = controlnet_block
+
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=mid_block_channel,
+                temb_channels=time_embed_dim,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                cross_attention_dim=cross_attention_dim,
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+            )
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=time_embed_dim,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+    @classmethod
+    def from_unet(
+        cls,
+        unet: UNet2DConditionModel,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+        load_weights_from_unet: bool = True,
+        conditioning_channels: int = 3,
+    ):
+        r"""
+        Instantiate a [`ControlNetModel`] from [`UNet2DConditionModel`].
+
+        Parameters:
+            unet (`UNet2DConditionModel`):
+                The UNet model weights to copy to the [`ControlNetModel`]. All configuration options are also copied
+                where applicable.
+        """
+        transformer_layers_per_block = (
+            unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1
+        )
+        encoder_hid_dim = unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None
+        encoder_hid_dim_type = unet.config.encoder_hid_dim_type if "encoder_hid_dim_type" in unet.config else None
+        addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None
+        addition_time_embed_dim = (
+            unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None
+        )
+
+        controlnet = cls(
+            encoder_hid_dim=encoder_hid_dim,
+            encoder_hid_dim_type=encoder_hid_dim_type,
+            addition_embed_type=addition_embed_type,
+            addition_time_embed_dim=addition_time_embed_dim,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=unet.config.in_channels,
+            flip_sin_to_cos=unet.config.flip_sin_to_cos,
+            freq_shift=unet.config.freq_shift,
+            down_block_types=unet.config.down_block_types,
+            only_cross_attention=unet.config.only_cross_attention,
+            block_out_channels=unet.config.block_out_channels,
+            layers_per_block=unet.config.layers_per_block,
+            downsample_padding=unet.config.downsample_padding,
+            mid_block_scale_factor=unet.config.mid_block_scale_factor,
+            act_fn=unet.config.act_fn,
+            norm_num_groups=unet.config.norm_num_groups,
+            norm_eps=unet.config.norm_eps,
+            cross_attention_dim=unet.config.cross_attention_dim,
+            attention_head_dim=unet.config.attention_head_dim,
+            num_attention_heads=unet.config.num_attention_heads,
+            use_linear_projection=unet.config.use_linear_projection,
+            class_embed_type=unet.config.class_embed_type,
+            num_class_embeds=unet.config.num_class_embeds,
+            upcast_attention=unet.config.upcast_attention,
+            resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
+            projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
+            mid_block_type=unet.config.mid_block_type,
+            controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
+            conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+            conditioning_channels=conditioning_channels,
+        )
+
+        if load_weights_from_unet:
+            controlnet.conv_in.load_state_dict(unet.conv_in.state_dict())
+            controlnet.ref_conv_in.load_state_dict(unet.conv_in.state_dict())
+            controlnet.time_proj.load_state_dict(unet.time_proj.state_dict())
+            controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
+
+            if controlnet.class_embedding:
+                controlnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
+
+            if hasattr(controlnet, "add_embedding"):
+                controlnet.add_embedding.load_state_dict(unet.add_embedding.state_dict())
+
+            controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict())
+            controlnet.mid_block.load_state_dict(unet.mid_block.state_dict())
+
+        return controlnet
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+    def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None:
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def process_encoder_hidden_states(
+        self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> torch.Tensor:
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds)
+            encoder_hidden_states = (encoder_hidden_states, image_embeds)
+        return encoder_hidden_states
+    
+    def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+        if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        controlnet_cond: torch.FloatTensor,
+        cat_dim: int = -2,
+        conditioning_scale: float = 1.0,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[AggregatorOutput, Tuple[Tuple[torch.FloatTensor, ...], torch.FloatTensor]]:
+        """
+        The [`Aggregator`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor.
+            timestep (`Union[torch.Tensor, float, int]`):
+                The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.Tensor`):
+                The encoder hidden states.
+            controlnet_cond (`torch.FloatTensor`):
+                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+            conditioning_scale (`float`, defaults to `1.0`):
+                The scale factor for ControlNet outputs.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
+                Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
+                timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
+                embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            added_cond_kwargs (`dict`):
+                Additional conditions for the Stable Diffusion XL UNet.
+            cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
+                A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+            return_dict (`bool`, defaults to `True`):
+                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.controlnet.ControlNetOutput`] **or** `tuple`:
+                If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
+                returned where the first element is the sample tensor.
+        """
+        # check channel order
+        channel_order = self.config.controlnet_conditioning_channel_order
+
+        if channel_order == "rgb":
+            # in rgb order by default
+            ...
+        else:
+            raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+
+        # timesteps does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        if self.config.addition_embed_type is not None:
+            if self.config.addition_embed_type == "text":
+                aug_emb = self.add_embedding(encoder_hidden_states)
+
+            elif self.config.addition_embed_type == "text_time":
+                if "text_embeds" not in added_cond_kwargs:
+                    raise ValueError(
+                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                    )
+                text_embeds = added_cond_kwargs.get("text_embeds")
+                if "time_ids" not in added_cond_kwargs:
+                    raise ValueError(
+                        f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                    )
+                time_ids = added_cond_kwargs.get("time_ids")
+                time_embeds = self.add_time_proj(time_ids.flatten())
+                time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+                add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+                add_embeds = add_embeds.to(emb.dtype)
+                aug_emb = self.add_embedding(add_embeds)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        encoder_hidden_states = self.process_encoder_hidden_states(
+            encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+
+        # 2. prepare input
+        cond_latent = self.conv_in(sample)
+        ref_latent = self.ref_conv_in(controlnet_cond)
+        batch_size, channel, height, width = cond_latent.shape
+        if self.pad_concat:
+            if cat_dim == -2 or cat_dim == 2:
+                concat_pad = torch.zeros(batch_size, channel, 1, width)
+            elif cat_dim == -1 or cat_dim == 3:
+                concat_pad = torch.zeros(batch_size, channel, height, 1)
+            else:
+                raise ValueError(f"Aggregator shall concat along spatial dimension, but is asked to concat dim: {cat_dim}.")
+            concat_pad = concat_pad.to(cond_latent.device, dtype=cond_latent.dtype)
+            sample = torch.cat([cond_latent, concat_pad, ref_latent], dim=cat_dim)
+        else:
+            sample = torch.cat([cond_latent, ref_latent], dim=cat_dim)
+
+        # 3. down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            sample, res_samples = downsample_block(
+                hidden_states=sample,
+                temb=emb,
+                cross_attention_kwargs=cross_attention_kwargs,
+            )
+
+            # rebuild sample: split and concat
+            if self.pad_concat:
+                batch_size, channel, height, width = sample.shape
+                if cat_dim == -2 or cat_dim == 2:
+                    cond_latent = sample[:, :, :height//2, :]
+                    ref_latent = sample[:, :, -(height//2):, :]
+                    concat_pad = torch.zeros(batch_size, channel, 1, width)
+                elif cat_dim == -1 or cat_dim == 3:
+                    cond_latent = sample[:, :, :, :width//2]
+                    ref_latent = sample[:, :, :, -(width//2):]
+                    concat_pad = torch.zeros(batch_size, channel, height, 1)
+                concat_pad = concat_pad.to(cond_latent.device, dtype=cond_latent.dtype)
+                sample = torch.cat([cond_latent, concat_pad, ref_latent], dim=cat_dim)
+                res_samples = res_samples[:-1] + (sample,)
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                cross_attention_kwargs=cross_attention_kwargs,
+            )
+
+        # 5. split samples and SFT.
+        controlnet_down_block_res_samples = ()
+        for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
+            batch_size, channel, height, width = down_block_res_sample.shape
+            if cat_dim == -2 or cat_dim == 2:
+                cond_latent = down_block_res_sample[:, :, :height//2, :]
+                ref_latent = down_block_res_sample[:, :, -(height//2):, :]
+            elif cat_dim == -1 or cat_dim == 3:
+                cond_latent = down_block_res_sample[:, :, :, :width//2]
+                ref_latent = down_block_res_sample[:, :, :, -(width//2):]
+            down_block_res_sample = controlnet_block((cond_latent, ref_latent), )
+            controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,)
+
+        down_block_res_samples = controlnet_down_block_res_samples
+
+        batch_size, channel, height, width = sample.shape
+        if cat_dim == -2 or cat_dim == 2:
+            cond_latent = sample[:, :, :height//2, :]
+            ref_latent = sample[:, :, -(height//2):, :]
+        elif cat_dim == -1 or cat_dim == 3:
+            cond_latent = sample[:, :, :, :width//2]
+            ref_latent = sample[:, :, :, -(width//2):]
+        mid_block_res_sample = self.controlnet_mid_block((cond_latent, ref_latent), )
+
+        # 6. scaling
+        down_block_res_samples = [sample*conditioning_scale for sample in down_block_res_samples]
+        mid_block_res_sample = mid_block_res_sample*conditioning_scale
+
+        if self.config.global_pool_conditions:
+            down_block_res_samples = [
+                torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples
+            ]
+            mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True)
+
+        if not return_dict:
+            return (down_block_res_samples, mid_block_res_sample)
+
+        return AggregatorOutput(
+            down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
+        )
+
+
+def zero_module(module):
+    for p in module.parameters():
+        nn.init.zeros_(p)
+    return module
diff --git a/modules/instantir/ip_adapter/__init__.py b/modules/instantir/ip_adapter/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/modules/instantir/ip_adapter/attention_processor.py b/modules/instantir/ip_adapter/attention_processor.py
new file mode 100644
index 000000000..ed6cf755f
--- /dev/null
+++ b/modules/instantir/ip_adapter/attention_processor.py
@@ -0,0 +1,1467 @@
+# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+class AdaLayerNorm(nn.Module):
+    def __init__(self, embedding_dim: int, time_embedding_dim: int = None):
+        super().__init__()
+
+        if time_embedding_dim is None:
+            time_embedding_dim = embedding_dim
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(time_embedding_dim, 2 * embedding_dim, bias=True)
+        nn.init.zeros_(self.linear.weight)
+        nn.init.zeros_(self.linear.bias)
+
+        self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
+
+    def forward(
+        self, x: torch.Tensor, timestep_embedding: torch.Tensor
+    ):
+        emb = self.linear(self.silu(timestep_embedding))
+        shift, scale = emb.view(len(x), 1, -1).chunk(2, dim=-1)
+        x = self.norm(x) * (1 + scale) + shift
+        return x
+
+
+class AttnProcessor(nn.Module):
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor(nn.Module):
+    r"""
+    Attention processor for IP-Adapater.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        ip_key = attn.head_to_batch_dim(ip_key)
+        ip_value = attn.head_to_batch_dim(ip_value)
+
+        ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+        ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+        ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class TA_IPAttnProcessor(nn.Module):
+    r"""
+    Attention processor for IP-Adapater.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, time_embedding_dim: int = None, scale=1.0, num_tokens=4):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+        self.ln_k_ip = AdaLayerNorm(hidden_size, time_embedding_dim)
+        self.ln_v_ip = AdaLayerNorm(hidden_size, time_embedding_dim)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        assert temb is not None, "Timestep embedding is needed for a time-aware attention processor."
+
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        # time-dependent adaLN
+        ip_key = self.ln_k_ip(ip_key, temb)
+        ip_value = self.ln_v_ip(ip_value, temb)
+
+        ip_key = attn.head_to_batch_dim(ip_key)
+        ip_value = attn.head_to_batch_dim(ip_value)
+
+        ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+        ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+        ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        external_kv=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        if external_kv:
+            key = torch.cat([key, external_kv.k], axis=1)
+            value = torch.cat([value, external_kv.v], axis=1)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class split_AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+        time_embedding_dim=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        external_kv=None,
+        temb=None,
+        cat_dim=-2,
+        original_shape=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            # 2d to sequence.
+            height, width = hidden_states.shape[-2:]
+            if cat_dim==-2 or cat_dim==2:
+                hidden_states_0 = hidden_states[:, :, :height//2, :]
+                hidden_states_1 = hidden_states[:, :, -(height//2):, :]
+            elif cat_dim==-1 or cat_dim==3:
+                hidden_states_0 = hidden_states[:, :, :, :width//2]
+                hidden_states_1 = hidden_states[:, :, :, -(width//2):]
+            batch_size, channel, height, width = hidden_states_0.shape
+            hidden_states_0 = hidden_states_0.view(batch_size, channel, height * width).transpose(1, 2)
+            hidden_states_1 = hidden_states_1.view(batch_size, channel, height * width).transpose(1, 2)
+        else:
+            # directly split sqeuence according to concat dim.
+            single_dim = original_shape[2] if cat_dim==-2 or cat_dim==2 else original_shape[1]
+            hidden_states_0 = hidden_states[:, :single_dim*single_dim,:]
+            hidden_states_1 = hidden_states[:, single_dim*(single_dim+1):,:]
+
+        hidden_states = torch.cat([hidden_states_0, hidden_states_1], dim=1)
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        if external_kv:
+            key = torch.cat([key, external_kv.k], dim=1)
+            value = torch.cat([value, external_kv.v], dim=1)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        # spatially split.
+        hidden_states_0, hidden_states_1 = hidden_states.chunk(2, dim=1)
+
+        if input_ndim == 4:
+            hidden_states_0 = hidden_states_0.transpose(-1, -2).reshape(batch_size, channel, height, width)
+            hidden_states_1 = hidden_states_1.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+            if cat_dim==-2 or cat_dim==2:
+                hidden_states_pad = torch.zeros(batch_size, channel, 1, width)
+            elif cat_dim==-1 or cat_dim==3:
+                hidden_states_pad = torch.zeros(batch_size, channel, height, 1)
+            hidden_states_pad = hidden_states_pad.to(hidden_states_0.device, dtype=hidden_states_0.dtype)
+            hidden_states = torch.cat([hidden_states_0, hidden_states_pad, hidden_states_1], dim=cat_dim)
+            assert hidden_states.shape == residual.shape, f"{hidden_states.shape} != {residual.shape}"
+        else:
+            batch_size, sequence_length, inner_dim = hidden_states.shape
+            hidden_states_pad = torch.zeros(batch_size, single_dim, inner_dim)
+            hidden_states_pad = hidden_states_pad.to(hidden_states_0.device, dtype=hidden_states_0.dtype)
+            hidden_states = torch.cat([hidden_states_0, hidden_states_pad, hidden_states_1], dim=1)
+            assert hidden_states.shape == residual.shape, f"{hidden_states.shape} != {residual.shape}"
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class sep_split_AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+        time_embedding_dim=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.ln_k_ref = AdaLayerNorm(hidden_size, time_embedding_dim)
+        self.ln_v_ref = AdaLayerNorm(hidden_size, time_embedding_dim)
+        # self.hidden_size = hidden_size
+        # self.cross_attention_dim = cross_attention_dim
+        # self.scale = scale
+        # self.num_tokens = num_tokens
+
+        # self.to_q_ref = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        # self.to_k_ref = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        # self.to_v_ref = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        external_kv=None,
+        temb=None,
+        cat_dim=-2,
+        original_shape=None,
+        ref_scale=1.0,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            # 2d to sequence.
+            height, width = hidden_states.shape[-2:]
+            if cat_dim==-2 or cat_dim==2:
+                hidden_states_0 = hidden_states[:, :, :height//2, :]
+                hidden_states_1 = hidden_states[:, :, -(height//2):, :]
+            elif cat_dim==-1 or cat_dim==3:
+                hidden_states_0 = hidden_states[:, :, :, :width//2]
+                hidden_states_1 = hidden_states[:, :, :, -(width//2):]
+            batch_size, channel, height, width = hidden_states_0.shape
+            hidden_states_0 = hidden_states_0.view(batch_size, channel, height * width).transpose(1, 2)
+            hidden_states_1 = hidden_states_1.view(batch_size, channel, height * width).transpose(1, 2)
+        else:
+            # directly split sqeuence according to concat dim.
+            single_dim = original_shape[2] if cat_dim==-2 or cat_dim==2 else original_shape[1]
+            hidden_states_0 = hidden_states[:, :single_dim*single_dim,:]
+            hidden_states_1 = hidden_states[:, single_dim*(single_dim+1):,:]
+
+        batch_size, sequence_length, _ = (
+            hidden_states_0.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states_0 = attn.group_norm(hidden_states_0.transpose(1, 2)).transpose(1, 2)
+            hidden_states_1 = attn.group_norm(hidden_states_1.transpose(1, 2)).transpose(1, 2)
+
+        query_0 = attn.to_q(hidden_states_0)
+        query_1 = attn.to_q(hidden_states_1)
+        key_0 = attn.to_k(hidden_states_0)
+        key_1 = attn.to_k(hidden_states_1)
+        value_0 = attn.to_v(hidden_states_0)
+        value_1 = attn.to_v(hidden_states_1)
+
+        # time-dependent adaLN
+        key_1 = self.ln_k_ref(key_1, temb)
+        value_1 = self.ln_v_ref(value_1, temb)
+
+        if external_kv:
+            key_1 = torch.cat([key_1, external_kv.k], dim=1)
+            value_1 = torch.cat([value_1, external_kv.v], dim=1)
+
+        inner_dim = key_0.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query_0 = query_0.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        query_1 = query_1.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key_0 = key_0.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        key_1 = key_1.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value_0 = value_0.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value_1 = value_1.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states_0 = F.scaled_dot_product_attention(
+            query_0, key_0, value_0, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states_1 = F.scaled_dot_product_attention(
+            query_1, key_1, value_1, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        # cross-attn
+        _hidden_states_0 = F.scaled_dot_product_attention(
+            query_0, key_1, value_1, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states_0 = hidden_states_0 + ref_scale * _hidden_states_0 * 10
+
+        # TODO: drop this cross-attn
+        _hidden_states_1 = F.scaled_dot_product_attention(
+            query_1, key_0, value_0, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+        hidden_states_1 = hidden_states_1 + ref_scale * _hidden_states_1
+
+        hidden_states_0 = hidden_states_0.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states_1 = hidden_states_1.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states_0 = hidden_states_0.to(query_0.dtype)
+        hidden_states_1 = hidden_states_1.to(query_1.dtype)
+
+
+        # linear proj
+        hidden_states_0 = attn.to_out[0](hidden_states_0)
+        hidden_states_1 = attn.to_out[0](hidden_states_1)
+        # dropout
+        hidden_states_0 = attn.to_out[1](hidden_states_0)
+        hidden_states_1 = attn.to_out[1](hidden_states_1)
+
+
+        if input_ndim == 4:
+            hidden_states_0 = hidden_states_0.transpose(-1, -2).reshape(batch_size, channel, height, width)
+            hidden_states_1 = hidden_states_1.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+            if cat_dim==-2 or cat_dim==2:
+                hidden_states_pad = torch.zeros(batch_size, channel, 1, width)
+            elif cat_dim==-1 or cat_dim==3:
+                hidden_states_pad = torch.zeros(batch_size, channel, height, 1)
+            hidden_states_pad = hidden_states_pad.to(hidden_states_0.device, dtype=hidden_states_0.dtype)
+            hidden_states = torch.cat([hidden_states_0, hidden_states_pad, hidden_states_1], dim=cat_dim)
+            assert hidden_states.shape == residual.shape, f"{hidden_states.shape} != {residual.shape}"
+        else:
+            batch_size, sequence_length, inner_dim = hidden_states.shape
+            hidden_states_pad = torch.zeros(batch_size, single_dim, inner_dim)
+            hidden_states_pad = hidden_states_pad.to(hidden_states_0.device, dtype=hidden_states_0.dtype)
+            hidden_states = torch.cat([hidden_states_0, hidden_states_pad, hidden_states_1], dim=1)
+            assert hidden_states.shape == residual.shape, f"{hidden_states.shape} != {residual.shape}"
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AdditiveKV_AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size: int = None,
+        cross_attention_dim: int = None,
+        time_embedding_dim: int = None,
+        additive_scale: float = 1.0,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.additive_scale = additive_scale
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        external_kv=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        assert temb is not None, "Timestep embedding is needed for a time-aware attention processor."
+        
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+
+        if external_kv:
+            key = external_kv.k
+            value = external_kv.v
+
+            key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            external_attn_output = F.scaled_dot_product_attention(
+                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            )
+
+            external_attn_output = external_attn_output.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+            hidden_states = hidden_states + self.additive_scale * external_attn_output
+
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class TA_AdditiveKV_AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size: int = None,
+        cross_attention_dim: int = None,
+        time_embedding_dim: int = None,
+        additive_scale: float = 1.0,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.ln_k = AdaLayerNorm(hidden_size, time_embedding_dim)
+        self.ln_v = AdaLayerNorm(hidden_size, time_embedding_dim)
+        self.additive_scale = additive_scale
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        external_kv=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        assert temb is not None, "Timestep embedding is needed for a time-aware attention processor."
+        
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+
+        if external_kv:
+            key = external_kv.k
+            value = external_kv.v
+
+            # time-dependent adaLN
+            key = self.ln_k(key, temb)
+            value = self.ln_v(value, temb)
+
+            key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            external_attn_output = F.scaled_dot_product_attention(
+                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            )
+
+            external_attn_output = external_attn_output.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+            hidden_states = hidden_states + self.additive_scale * external_attn_output
+
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        if isinstance(encoder_hidden_states, tuple):
+            # FIXME: now hard coded to single image prompt.
+            batch_size, _, hid_dim = encoder_hidden_states[0].shape
+            ip_tokens = encoder_hidden_states[1][0]
+            encoder_hidden_states = torch.cat([encoder_hidden_states[0], ip_tokens], dim=1)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+
+        ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        ip_hidden_states = F.scaled_dot_product_attention(
+            query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+        )
+
+        ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        ip_hidden_states = ip_hidden_states.to(query.dtype)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class TA_IPAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, time_embedding_dim: int = None, scale=1.0, num_tokens=4):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.ln_k_ip = AdaLayerNorm(hidden_size, time_embedding_dim)
+        self.ln_v_ip = AdaLayerNorm(hidden_size, time_embedding_dim)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        external_kv=None,
+        temb=None,
+    ):
+        assert temb is not None, "Timestep embedding is needed for a time-aware attention processor."
+
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        if not isinstance(encoder_hidden_states, tuple):
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+        else:
+            # FIXME: now hard coded to single image prompt.
+            batch_size, _, hid_dim = encoder_hidden_states[0].shape
+            ip_hidden_states = encoder_hidden_states[1][0]
+            encoder_hidden_states = encoder_hidden_states[0]
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        if external_kv:
+            key = torch.cat([key, external_kv.k], axis=1)
+            value = torch.cat([value, external_kv.v], axis=1)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # for ip-adapter
+        ip_key = self.to_k_ip(ip_hidden_states)
+        ip_value = self.to_v_ip(ip_hidden_states)
+        
+        # time-dependent adaLN
+        ip_key = self.ln_k_ip(ip_key, temb)
+        ip_value = self.ln_v_ip(ip_value, temb)
+
+        ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        ip_hidden_states = F.scaled_dot_product_attention(
+            query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+        )
+
+        ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        ip_hidden_states = ip_hidden_states.to(query.dtype)
+
+        hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+## for controlnet
+class CNAttnProcessor:
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(self, num_tokens=4):
+        self.num_tokens = num_tokens
+
+    def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class CNAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self, num_tokens=4):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.num_tokens = num_tokens
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+def init_attn_proc(unet, ip_adapter_tokens=16, use_lcm=False, use_adaln=True, use_external_kv=False):
+    attn_procs = {}
+    unet_sd = unet.state_dict()
+    for name in unet.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+        if cross_attention_dim is None:
+            if use_external_kv:
+                attn_procs[name] = AdditiveKV_AttnProcessor2_0(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                    time_embedding_dim=1280,
+                ) if hasattr(F, "scaled_dot_product_attention") else AdditiveKV_AttnProcessor()
+            else:
+                attn_procs[name] = AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnProcessor()
+        else:
+            if use_adaln:
+                layer_name = name.split(".processor")[0]
+                if use_lcm:
+                    weights = {
+                        "to_k_ip.weight": unet_sd[layer_name + ".to_k.base_layer.weight"],
+                        "to_v_ip.weight": unet_sd[layer_name + ".to_v.base_layer.weight"],
+                    }
+                else:
+                    weights = {
+                        "to_k_ip.weight": unet_sd[layer_name + ".to_k.weight"],
+                        "to_v_ip.weight": unet_sd[layer_name + ".to_v.weight"],
+                    }
+                attn_procs[name] = TA_IPAttnProcessor2_0(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        num_tokens=ip_adapter_tokens,
+                        time_embedding_dim=1280,
+                    ) if hasattr(F, "scaled_dot_product_attention") else \
+                    TA_IPAttnProcessor(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        num_tokens=ip_adapter_tokens,
+                        time_embedding_dim=1280,
+                    )
+                attn_procs[name].load_state_dict(weights, strict=False)
+            else:
+                attn_procs[name] = AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnProcessor()
+
+    return attn_procs
+
+
+def init_aggregator_attn_proc(unet, use_adaln=False, split_attn=False):
+    attn_procs = {}
+    unet_sd = unet.state_dict()
+    for name in unet.attn_processors.keys():
+        # get layer name and hidden dim
+        cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = unet.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = unet.config.block_out_channels[block_id]
+        # init attn proc
+        if split_attn:
+            # layer_name = name.split(".processor")[0]
+            # weights = {
+            #     "to_q_ref.weight": unet_sd[layer_name + ".to_q.weight"],
+            #     "to_k_ref.weight": unet_sd[layer_name + ".to_k.weight"],
+            #     "to_v_ref.weight": unet_sd[layer_name + ".to_v.weight"],
+            # }
+            attn_procs[name] = (
+                sep_split_AttnProcessor2_0(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=hidden_size,
+                    time_embedding_dim=1280,
+                )
+                if use_adaln
+                else split_AttnProcessor2_0(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=cross_attention_dim,
+                    time_embedding_dim=1280,
+                )
+            )
+            # attn_procs[name].load_state_dict(weights, strict=False)
+        else:
+            attn_procs[name] = (
+                AttnProcessor2_0(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=hidden_size,
+                )
+                if hasattr(F, "scaled_dot_product_attention")
+                else AttnProcessor(
+                    hidden_size=hidden_size,
+                    cross_attention_dim=hidden_size,
+                )
+            )
+
+    return attn_procs
diff --git a/modules/instantir/ip_adapter/ip_adapter.py b/modules/instantir/ip_adapter/ip_adapter.py
new file mode 100644
index 000000000..10f01d4f3
--- /dev/null
+++ b/modules/instantir/ip_adapter/ip_adapter.py
@@ -0,0 +1,236 @@
+import os
+import torch
+from typing import List
+from collections import namedtuple, OrderedDict
+
+def is_torch2_available():
+    return hasattr(torch.nn.functional, "scaled_dot_product_attention")
+
+if is_torch2_available():
+    from .attention_processor import (
+        AttnProcessor2_0 as AttnProcessor,
+    )
+    from .attention_processor import (
+        CNAttnProcessor2_0 as CNAttnProcessor,
+    )
+    from .attention_processor import (
+        IPAttnProcessor2_0 as IPAttnProcessor,
+    )
+    from .attention_processor import (
+        TA_IPAttnProcessor2_0 as TA_IPAttnProcessor,
+    )
+else:
+    from .attention_processor import AttnProcessor, CNAttnProcessor, IPAttnProcessor, TA_IPAttnProcessor
+
+
+class ImageProjModel(torch.nn.Module):
+    """Projection Model"""
+
+    def __init__(self, cross_attention_dim=2048, clip_embeddings_dim=1280, clip_extra_context_tokens=4):
+        super().__init__()
+
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+class MLPProjModel(torch.nn.Module):
+    """SD model with image prompt"""
+    def __init__(self, cross_attention_dim=2048, clip_embeddings_dim=1280):
+        super().__init__()
+        
+        self.proj = torch.nn.Sequential(
+            torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim),
+            torch.nn.GELU(),
+            torch.nn.Linear(clip_embeddings_dim, cross_attention_dim),
+            torch.nn.LayerNorm(cross_attention_dim)
+        )
+        
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class MultiIPAdapterImageProjection(torch.nn.Module):
+    def __init__(self, IPAdapterImageProjectionLayers):
+        super().__init__()
+        self.image_projection_layers = torch.nn.ModuleList(IPAdapterImageProjectionLayers)
+
+    def forward(self, image_embeds: List[torch.FloatTensor]):
+        projected_image_embeds = []
+
+        # currently, we accept `image_embeds` as
+        #  1. a tensor (deprecated) with shape [batch_size, embed_dim] or [batch_size, sequence_length, embed_dim]
+        #  2. list of `n` tensors where `n` is number of ip-adapters, each tensor can hae shape [batch_size, num_images, embed_dim] or [batch_size, num_images, sequence_length, embed_dim]
+        if not isinstance(image_embeds, list):
+            image_embeds = [image_embeds.unsqueeze(1)]
+
+        if len(image_embeds) != len(self.image_projection_layers):
+            raise ValueError(
+                f"image_embeds must have the same length as image_projection_layers, got {len(image_embeds)} and {len(self.image_projection_layers)}"
+            )
+
+        for image_embed, image_projection_layer in zip(image_embeds, self.image_projection_layers):
+            batch_size, num_images = image_embed.shape[0], image_embed.shape[1]
+            image_embed = image_embed.reshape((batch_size * num_images,) + image_embed.shape[2:])
+            image_embed = image_projection_layer(image_embed)
+            # image_embed = image_embed.reshape((batch_size, num_images) + image_embed.shape[1:])
+
+            projected_image_embeds.append(image_embed)
+
+        return projected_image_embeds
+
+
+class IPAdapter(torch.nn.Module):
+    """IP-Adapter"""
+    def __init__(self, unet, image_proj_model, adapter_modules, ckpt_path=None):
+        super().__init__()
+        self.unet = unet
+        self.image_proj = image_proj_model
+        self.ip_adapter = adapter_modules
+
+        if ckpt_path is not None:
+            self.load_from_checkpoint(ckpt_path)
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, image_embeds):
+        ip_tokens = self.image_proj(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states).sample
+        return noise_pred
+
+    def load_from_checkpoint(self, ckpt_path: str):
+        # Calculate original checksums
+        orig_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj.parameters()]))
+        orig_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.ip_adapter.parameters()]))
+
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+        keys = list(state_dict.keys())
+        if keys != ["image_proj", "ip_adapter"]:
+            state_dict = revise_state_dict(state_dict)
+
+        # Load state dict for image_proj_model and adapter_modules
+        self.image_proj.load_state_dict(state_dict["image_proj"], strict=True)
+        self.ip_adapter.load_state_dict(state_dict["ip_adapter"], strict=True)
+
+        # Calculate new checksums
+        new_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj.parameters()]))
+        new_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.ip_adapter.parameters()]))
+
+        # Verify if the weights have changed
+        assert orig_ip_proj_sum != new_ip_proj_sum, "Weights of image_proj_model did not change!"
+        assert orig_adapter_sum != new_adapter_sum, "Weights of adapter_modules did not change!"
+
+
+class IPAdapterPlus(torch.nn.Module):
+    """IP-Adapter"""
+    def __init__(self, unet, image_proj_model, adapter_modules, ckpt_path=None):
+        super().__init__()
+        self.unet = unet
+        self.image_proj = image_proj_model
+        self.ip_adapter = adapter_modules
+
+        if ckpt_path is not None:
+            self.load_from_checkpoint(ckpt_path)
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, image_embeds):
+        ip_tokens = self.image_proj(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states).sample
+        return noise_pred
+
+    def load_from_checkpoint(self, ckpt_path: str):
+        # Calculate original checksums
+        orig_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj.parameters()]))
+        orig_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.ip_adapter.parameters()]))
+        org_unet_sum = []
+        for attn_name, attn_proc in self.unet.attn_processors.items():
+            if isinstance(attn_proc, (TA_IPAttnProcessor, IPAttnProcessor)):
+                org_unet_sum.append(torch.sum(torch.stack([torch.sum(p) for p in attn_proc.parameters()])))
+        org_unet_sum = torch.sum(torch.stack(org_unet_sum))
+
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+        keys = list(state_dict.keys())
+        if keys != ["image_proj", "ip_adapter"]:
+            state_dict = revise_state_dict(state_dict)
+
+        # Check if 'latents' exists in both the saved state_dict and the current model's state_dict
+        strict_load_image_proj_model = True
+        if "latents" in state_dict["image_proj"] and "latents" in self.image_proj.state_dict():
+            # Check if the shapes are mismatched
+            if state_dict["image_proj"]["latents"].shape != self.image_proj.state_dict()["latents"].shape:
+                print(f"Shapes of 'image_proj.latents' in checkpoint {ckpt_path} and current model do not match.")
+                print("Removing 'latents' from checkpoint and loading the rest of the weights.")
+                del state_dict["image_proj"]["latents"]
+                strict_load_image_proj_model = False
+
+        # Load state dict for image_proj_model and adapter_modules
+        self.image_proj.load_state_dict(state_dict["image_proj"], strict=strict_load_image_proj_model)
+        missing_key, unexpected_key = self.ip_adapter.load_state_dict(state_dict["ip_adapter"], strict=False)
+        if len(missing_key) > 0:
+            for ms in missing_key:
+                if "ln" not in ms:
+                    raise ValueError(f"Missing key in adapter_modules: {len(missing_key)}")
+        if len(unexpected_key) > 0:
+            raise ValueError(f"Unexpected key in adapter_modules: {len(unexpected_key)}")
+
+        # Calculate new checksums
+        new_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj.parameters()]))
+        new_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.ip_adapter.parameters()]))
+
+        # Verify if the weights loaded to unet
+        unet_sum = []
+        for attn_name, attn_proc in self.unet.attn_processors.items():
+            if isinstance(attn_proc, (TA_IPAttnProcessor, IPAttnProcessor)):
+                unet_sum.append(torch.sum(torch.stack([torch.sum(p) for p in attn_proc.parameters()])))
+        unet_sum = torch.sum(torch.stack(unet_sum))
+
+        assert org_unet_sum != unet_sum, "Weights of adapter_modules in unet did not change!"
+        assert (unet_sum - new_adapter_sum < 1e-4), "Weights of adapter_modules did not load to unet!"
+
+        # Verify if the weights have changed
+        assert orig_ip_proj_sum != new_ip_proj_sum, "Weights of image_proj_model did not change!"
+        assert orig_adapter_sum != new_adapter_sum, "Weights of adapter_mod`ules did not change!"
+
+
+class IPAdapterXL(IPAdapter):
+    """SDXL"""
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, unet_added_cond_kwargs, image_embeds):
+        ip_tokens = self.image_proj(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states, added_cond_kwargs=unet_added_cond_kwargs).sample
+        return noise_pred
+
+
+class IPAdapterPlusXL(IPAdapterPlus):
+    """IP-Adapter with fine-grained features"""
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, unet_added_cond_kwargs, image_embeds):
+        ip_tokens = self.image_proj(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states, added_cond_kwargs=unet_added_cond_kwargs).sample
+        return noise_pred
+
+
+class IPAdapterFull(IPAdapterPlus):
+    """IP-Adapter with full features"""
+
+    def init_proj(self):
+        image_proj_model = MLPProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.hidden_size,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
diff --git a/modules/instantir/ip_adapter/resampler.py b/modules/instantir/ip_adapter/resampler.py
new file mode 100644
index 000000000..72295f90b
--- /dev/null
+++ b/modules/instantir/ip_adapter/resampler.py
@@ -0,0 +1,158 @@
+# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py
+
+import math
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1280,
+        depth=4,
+        dim_head=64,
+        heads=20,
+        num_queries=64,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+        max_seq_len: int = 257,  # CLIP tokens + CLS token
+        apply_pos_emb: bool = False,
+        num_latents_mean_pooled: int = 0,  # number of latents derived from mean pooled representation of the sequence
+    ):
+        super().__init__()
+        self.pos_emb = nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.to_latents_from_mean_pooled_seq = (
+            nn.Sequential(
+                nn.LayerNorm(dim),
+                nn.Linear(dim, dim * num_latents_mean_pooled),
+                Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled),
+            )
+            if num_latents_mean_pooled > 0
+            else None
+        )
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        if self.pos_emb is not None:
+            n, device = x.shape[1], x.device
+            pos_emb = self.pos_emb(torch.arange(n, device=device))
+            x = x + pos_emb
+
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        if self.to_latents_from_mean_pooled_seq:
+            meanpooled_seq = masked_mean(x, dim=1, mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool))
+            meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq)
+            latents = torch.cat((meanpooled_latents, latents), dim=-2)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
+def masked_mean(t, *, dim, mask=None):
+    if mask is None:
+        return t.mean(dim=dim)
+
+    denom = mask.sum(dim=dim, keepdim=True)
+    mask = rearrange(mask, "b n -> b n 1")
+    masked_t = t.masked_fill(~mask, 0.0)
+
+    return masked_t.sum(dim=dim) / denom.clamp(min=1e-5)
diff --git a/modules/instantir/ip_adapter/utils.py b/modules/instantir/ip_adapter/utils.py
new file mode 100644
index 000000000..64c45cd85
--- /dev/null
+++ b/modules/instantir/ip_adapter/utils.py
@@ -0,0 +1,248 @@
+import torch
+from collections import namedtuple, OrderedDict
+from safetensors import safe_open
+from .attention_processor import init_attn_proc
+from .ip_adapter import MultiIPAdapterImageProjection
+from .resampler import Resampler
+from transformers import (
+    AutoModel, AutoImageProcessor,
+    CLIPVisionModelWithProjection, CLIPImageProcessor)
+
+
+def init_adapter_in_unet(
+        unet,
+        image_proj_model=None,
+        pretrained_model_path_or_dict=None,
+        adapter_tokens=64,
+        embedding_dim=None,
+        use_lcm=False,
+        use_adaln=True,
+    ):
+        device = unet.device
+        dtype = unet.dtype
+        if image_proj_model is None:
+            assert embedding_dim is not None, "embedding_dim must be provided if image_proj_model is None."
+            image_proj_model = Resampler(
+                embedding_dim=embedding_dim,
+                output_dim=unet.config.cross_attention_dim,
+                num_queries=adapter_tokens,
+            )
+        if pretrained_model_path_or_dict is not None:
+            if not isinstance(pretrained_model_path_or_dict, dict):
+                if pretrained_model_path_or_dict.endswith(".safetensors"):
+                    state_dict = {"image_proj": {}, "ip_adapter": {}}
+                    with safe_open(pretrained_model_path_or_dict, framework="pt", device=unet.device) as f:
+                        for key in f.keys():
+                            if key.startswith("image_proj."):
+                                state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                            elif key.startswith("ip_adapter."):
+                                state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+                else:
+                    state_dict = torch.load(pretrained_model_path_or_dict, map_location=unet.device)
+            else:
+                state_dict = pretrained_model_path_or_dict
+            keys = list(state_dict.keys())
+            if "image_proj" not in keys and "ip_adapter" not in keys:
+                state_dict = revise_state_dict(state_dict)
+
+        # Creat IP cross-attention in unet.
+        attn_procs = init_attn_proc(unet, adapter_tokens, use_lcm, use_adaln)
+        unet.set_attn_processor(attn_procs)
+
+        # Load pretrinaed model if needed.
+        if pretrained_model_path_or_dict is not None:
+            if "ip_adapter" in state_dict.keys():
+                adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+                missing, unexpected = adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=False)
+                for mk in missing:
+                    if "ln" not in mk:
+                        raise ValueError(f"Missing keys in adapter_modules: {missing}")
+            if "image_proj" in state_dict.keys():
+                image_proj_model.load_state_dict(state_dict["image_proj"])
+
+        # Load image projectors into iterable ModuleList.
+        image_projection_layers = []
+        image_projection_layers.append(image_proj_model)
+        unet.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
+
+        # Adjust unet config to handle addtional ip hidden states.
+        unet.config.encoder_hid_dim_type = "ip_image_proj"
+        unet.to(dtype=dtype, device=device)
+
+
+def load_adapter_to_pipe(
+        pipe,
+        pretrained_model_path_or_dict,
+        image_encoder_or_path=None,
+        feature_extractor_or_path=None,
+        use_clip_encoder=False,
+        adapter_tokens=64,
+        use_lcm=False,
+        use_adaln=True,
+    ):
+
+        if not isinstance(pretrained_model_path_or_dict, dict):
+            if pretrained_model_path_or_dict.endswith(".safetensors"):
+                state_dict = {"image_proj": {}, "ip_adapter": {}}
+                with safe_open(pretrained_model_path_or_dict, framework="pt", device=pipe.device) as f:
+                    for key in f.keys():
+                        if key.startswith("image_proj."):
+                            state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                        elif key.startswith("ip_adapter."):
+                            state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+            else:
+                state_dict = torch.load(pretrained_model_path_or_dict, map_location=pipe.device)
+        else:
+            state_dict = pretrained_model_path_or_dict
+        keys = list(state_dict.keys())
+        if "image_proj" not in keys and "ip_adapter" not in keys:
+            state_dict = revise_state_dict(state_dict)
+
+        # load CLIP image encoder here if it has not been registered to the pipeline yet
+        if image_encoder_or_path is not None:
+            if isinstance(image_encoder_or_path, str):
+                feature_extractor_or_path = image_encoder_or_path if feature_extractor_or_path is None else feature_extractor_or_path
+
+                image_encoder_or_path = (
+                    CLIPVisionModelWithProjection.from_pretrained(
+                        image_encoder_or_path
+                    ) if use_clip_encoder else
+                    AutoModel.from_pretrained(image_encoder_or_path)
+                )
+
+        if feature_extractor_or_path is not None:
+            if isinstance(feature_extractor_or_path, str):
+                feature_extractor_or_path = (
+                    CLIPImageProcessor() if use_clip_encoder else
+                    AutoImageProcessor.from_pretrained(feature_extractor_or_path)
+                )
+
+        # create image encoder if it has not been registered to the pipeline yet
+        if hasattr(pipe, "image_encoder") and getattr(pipe, "image_encoder", None) is None:
+            image_encoder = image_encoder_or_path.to(pipe.device, dtype=pipe.dtype)
+            pipe.register_modules(image_encoder=image_encoder)
+        else:
+            image_encoder = pipe.image_encoder
+
+        # create feature extractor if it has not been registered to the pipeline yet
+        if hasattr(pipe, "feature_extractor") and getattr(pipe, "feature_extractor", None) is None:
+            feature_extractor = feature_extractor_or_path
+            pipe.register_modules(feature_extractor=feature_extractor)
+        else:
+            feature_extractor = pipe.feature_extractor
+
+        # load adapter into unet
+        unet = getattr(pipe, pipe.unet_name) if not hasattr(pipe, "unet") else pipe.unet
+        attn_procs = init_attn_proc(unet, adapter_tokens, use_lcm, use_adaln)
+        unet.set_attn_processor(attn_procs)
+        image_proj_model = Resampler(
+            embedding_dim=image_encoder.config.hidden_size,
+            output_dim=unet.config.cross_attention_dim,
+            num_queries=adapter_tokens,
+        )
+
+        # Load pretrinaed model if needed.
+        if "ip_adapter" in state_dict.keys():
+            adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+            missing, unexpected = adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=False)
+            for mk in missing:
+                if "ln" not in mk:
+                    raise ValueError(f"Missing keys in adapter_modules: {missing}")
+        if "image_proj" in state_dict.keys():
+            image_proj_model.load_state_dict(state_dict["image_proj"])
+
+        # convert IP-Adapter Image Projection layers to diffusers
+        image_projection_layers = []
+        image_projection_layers.append(image_proj_model)
+        unet.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
+
+        # Adjust unet config to handle addtional ip hidden states.
+        unet.config.encoder_hid_dim_type = "ip_image_proj"
+        unet.to(dtype=pipe.dtype, device=pipe.device)
+
+
+def revise_state_dict(old_state_dict_or_path, map_location="cpu"):
+    new_state_dict = OrderedDict()
+    new_state_dict["image_proj"] = OrderedDict()
+    new_state_dict["ip_adapter"] = OrderedDict()
+    if isinstance(old_state_dict_or_path, str):
+        old_state_dict = torch.load(old_state_dict_or_path, map_location=map_location)
+    else:
+        old_state_dict = old_state_dict_or_path
+    for name, weight in old_state_dict.items():
+        if name.startswith("image_proj_model."):
+            new_state_dict["image_proj"][name[len("image_proj_model."):]] = weight
+        elif name.startswith("adapter_modules."):
+            new_state_dict["ip_adapter"][name[len("adapter_modules."):]] = weight
+    return new_state_dict
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+def encode_image(image_encoder, feature_extractor, image, device, num_images_per_prompt, output_hidden_states=None):
+    dtype = next(image_encoder.parameters()).dtype
+
+    if not isinstance(image, torch.Tensor):
+        image = feature_extractor(image, return_tensors="pt").pixel_values
+
+    image = image.to(device=device, dtype=dtype)
+    if output_hidden_states:
+        image_enc_hidden_states = image_encoder(image, output_hidden_states=True).hidden_states[-2]
+        image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+        return image_enc_hidden_states
+    else:
+        if isinstance(image_encoder, CLIPVisionModelWithProjection):
+            # CLIP image encoder.
+            image_embeds = image_encoder(image).image_embeds
+        else:
+            # DINO image encoder.
+            image_embeds = image_encoder(image).last_hidden_state
+        image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+        return image_embeds
+
+
+def prepare_training_image_embeds(
+    image_encoder, feature_extractor,
+    ip_adapter_image, ip_adapter_image_embeds,
+    device, drop_rate, output_hidden_state, idx_to_replace=None
+):
+    if ip_adapter_image_embeds is None:
+        if not isinstance(ip_adapter_image, list):
+            ip_adapter_image = [ip_adapter_image]
+
+        # if len(ip_adapter_image) != len(unet.encoder_hid_proj.image_projection_layers):
+        #     raise ValueError(
+        #         f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+        #     )
+
+        image_embeds = []
+        for single_ip_adapter_image in ip_adapter_image:
+            if idx_to_replace is None:
+                idx_to_replace = torch.rand(len(single_ip_adapter_image)) < drop_rate
+            zero_ip_adapter_image = torch.zeros_like(single_ip_adapter_image)
+            single_ip_adapter_image[idx_to_replace] = zero_ip_adapter_image[idx_to_replace]
+            single_image_embeds = encode_image(
+                image_encoder, feature_extractor, single_ip_adapter_image, device, 1, output_hidden_state
+            )
+            single_image_embeds = torch.stack([single_image_embeds], dim=1) # FIXME
+
+            image_embeds.append(single_image_embeds)
+    else:
+        repeat_dims = [1]
+        image_embeds = []
+        for single_image_embeds in ip_adapter_image_embeds:
+            if do_classifier_free_guidance:
+                single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+                single_image_embeds = single_image_embeds.repeat(
+                    num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                )
+                single_negative_image_embeds = single_negative_image_embeds.repeat(
+                    num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+                )
+                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+            else:
+                single_image_embeds = single_image_embeds.repeat(
+                    num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                )
+            image_embeds.append(single_image_embeds)
+
+    return image_embeds
\ No newline at end of file
diff --git a/modules/instantir/lcm_single_step_scheduler.py b/modules/instantir/lcm_single_step_scheduler.py
new file mode 100644
index 000000000..a32affdc2
--- /dev/null
+++ b/modules/instantir/lcm_single_step_scheduler.py
@@ -0,0 +1,537 @@
+# Copyright 2023 Stanford University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import BaseOutput, logging
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class LCMSingleStepSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+            The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+            `pred_original_sample` can be used to preview progress or for guidance.
+    """
+
+    denoised: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+    (1-beta) over time from t = [0,1].
+
+    Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+    to that part of the diffusion process.
+
+
+    Args:
+        num_diffusion_timesteps (`int`): the number of betas to produce.
+        max_beta (`float`): the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
+
+    Returns:
+        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+    """
+    if alpha_transform_type == "cosine":
+
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+    return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas: torch.FloatTensor) -> torch.FloatTensor:
+    """
+    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+    Args:
+        betas (`torch.FloatTensor`):
+            the betas that the scheduler is being initialized with.
+
+    Returns:
+        `torch.FloatTensor`: rescaled betas with zero terminal SNR
+    """
+    # Convert betas to alphas_bar_sqrt
+    alphas = 1.0 - betas
+    alphas_cumprod = torch.cumprod(alphas, dim=0)
+    alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+    # Store old values.
+    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+    # Shift so the last timestep is zero.
+    alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+    # Scale so the first timestep is back to the old value.
+    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+    # Convert alphas_bar_sqrt to betas
+    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
+    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
+    alphas = torch.cat([alphas_bar[0:1], alphas])
+    betas = 1 - alphas
+
+    return betas
+
+
+class LCMSingleStepScheduler(SchedulerMixin, ConfigMixin):
+    """
+    `LCMSingleStepScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with
+    non-Markovian guidance.
+
+    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~ConfigMixin`] takes care of storing all config
+    attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be
+    accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving
+    functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions.
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            The number of diffusion steps to train the model.
+        beta_start (`float`, defaults to 0.0001):
+            The starting `beta` value of inference.
+        beta_end (`float`, defaults to 0.02):
+            The final `beta` value.
+        beta_schedule (`str`, defaults to `"linear"`):
+            The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+        trained_betas (`np.ndarray`, *optional*):
+            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+        original_inference_steps (`int`, *optional*, defaults to 50):
+            The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we
+            will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule.
+        clip_sample (`bool`, defaults to `True`):
+            Clip the predicted sample for numerical stability.
+        clip_sample_range (`float`, defaults to 1.0):
+            The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+        set_alpha_to_one (`bool`, defaults to `True`):
+            Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
+            there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+            otherwise it uses the alpha value at step 0.
+        steps_offset (`int`, defaults to 0):
+            An offset added to the inference steps. You can use a combination of `offset=1` and
+            `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+            Diffusion.
+        prediction_type (`str`, defaults to `epsilon`, *optional*):
+            Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+            `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+            Video](https://imagen.research.google/video/paper.pdf) paper).
+        thresholding (`bool`, defaults to `False`):
+            Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+            as Stable Diffusion.
+        dynamic_thresholding_ratio (`float`, defaults to 0.995):
+            The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+        sample_max_value (`float`, defaults to 1.0):
+            The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+        timestep_spacing (`str`, defaults to `"leading"`):
+            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+        timestep_scaling (`float`, defaults to 10.0):
+            The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions
+            `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation
+            error at the default of `10.0` is already pretty small).
+        rescale_betas_zero_snr (`bool`, defaults to `False`):
+            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+            dark samples instead of limiting it to samples with medium brightness. Loosely related to
+            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+    """
+
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        beta_start: float = 0.00085,
+        beta_end: float = 0.012,
+        beta_schedule: str = "scaled_linear",
+        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+        original_inference_steps: int = 50,
+        clip_sample: bool = False,
+        clip_sample_range: float = 1.0,
+        set_alpha_to_one: bool = True,
+        steps_offset: int = 0,
+        prediction_type: str = "epsilon",
+        thresholding: bool = False,
+        dynamic_thresholding_ratio: float = 0.995,
+        sample_max_value: float = 1.0,
+        timestep_spacing: str = "leading",
+        timestep_scaling: float = 10.0,
+        rescale_betas_zero_snr: bool = False,
+    ):
+        if trained_betas is not None:
+            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+        elif beta_schedule == "linear":
+            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+        elif beta_schedule == "scaled_linear":
+            # this schedule is very specific to the latent diffusion model.
+            self.betas = (
+                torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+            )
+        elif beta_schedule == "squaredcos_cap_v2":
+            # Glide cosine schedule
+            self.betas = betas_for_alpha_bar(num_train_timesteps)
+        else:
+            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+        # Rescale for zero SNR
+        if rescale_betas_zero_snr:
+            self.betas = rescale_zero_terminal_snr(self.betas)
+
+        self.alphas = 1.0 - self.betas
+        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+        # At every step in ddim, we are looking into the previous alphas_cumprod
+        # For the final step, there is no previous alphas_cumprod because we are already at 0
+        # `set_alpha_to_one` decides whether we set this parameter simply to one or
+        # whether we use the final alpha of the "non-previous" one.
+        self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+        # standard deviation of the initial noise distribution
+        self.init_noise_sigma = 1.0
+
+        # setable values
+        self.num_inference_steps = None
+        self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
+
+        self._step_index = None
+
+    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+    def _init_step_index(self, timestep):
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.to(self.timesteps.device)
+
+        index_candidates = (self.timesteps == timestep).nonzero()
+
+        # The sigma index that is taken for the **very** first `step`
+        # is always the second index (or the last index if there is only 1)
+        # This way we can ensure we don't accidentally skip a sigma in
+        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+        if len(index_candidates) > 1:
+            step_index = index_candidates[1]
+        else:
+            step_index = index_candidates[0]
+
+        self._step_index = step_index.item()
+
+    @property
+    def step_index(self):
+        return self._step_index
+
+    def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+        """
+        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+        current timestep.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The input sample.
+            timestep (`int`, *optional*):
+                The current timestep in the diffusion chain.
+        Returns:
+            `torch.FloatTensor`:
+                A scaled input sample.
+        """
+        return sample
+
+    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+    def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+        """
+        "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+        prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+        s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+        pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+        photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+        https://arxiv.org/abs/2205.11487
+        """
+        dtype = sample.dtype
+        batch_size, channels, *remaining_dims = sample.shape
+
+        if dtype not in (torch.float32, torch.float64):
+            sample = sample.float()  # upcast for quantile calculation, and clamp not implemented for cpu half
+
+        # Flatten sample for doing quantile calculation along each image
+        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+        abs_sample = sample.abs()  # "a certain percentile absolute pixel value"
+
+        s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+        s = torch.clamp(
+            s, min=1, max=self.config.sample_max_value
+        )  # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+        s = s.unsqueeze(1)  # (batch_size, 1) because clamp will broadcast along dim=0
+        sample = torch.clamp(sample, -s, s) / s  # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+        sample = sample.reshape(batch_size, channels, *remaining_dims)
+        sample = sample.to(dtype)
+
+        return sample
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int = None,
+        device: Union[str, torch.device] = None,
+        original_inference_steps: Optional[int] = None,
+        strength: int = 1.0,
+        timesteps: Optional[list] = None,
+    ):
+        """
+        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+        Args:
+            num_inference_steps (`int`):
+                The number of diffusion steps used when generating samples with a pre-trained model.
+            device (`str` or `torch.device`, *optional*):
+                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+            original_inference_steps (`int`, *optional*):
+                The original number of inference steps, which will be used to generate a linearly-spaced timestep
+                schedule (which is different from the standard `diffusers` implementation). We will then take
+                `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as
+                our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute.
+        """
+
+        if num_inference_steps is not None and timesteps is not None:
+            raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.")
+
+        if timesteps is not None:
+            for i in range(1, len(timesteps)):
+                if timesteps[i] >= timesteps[i - 1]:
+                    raise ValueError("`custom_timesteps` must be in descending order.")
+
+            if timesteps[0] >= self.config.num_train_timesteps:
+                raise ValueError(
+                    f"`timesteps` must start before `self.config.train_timesteps`:"
+                    f" {self.config.num_train_timesteps}."
+                )
+
+            timesteps = np.array(timesteps, dtype=np.int64)
+        else:
+            if num_inference_steps > self.config.num_train_timesteps:
+                raise ValueError(
+                    f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+                    f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+                    f" maximal {self.config.num_train_timesteps} timesteps."
+                )
+
+            self.num_inference_steps = num_inference_steps
+            original_steps = (
+                original_inference_steps if original_inference_steps is not None else self.config.original_inference_steps
+            )
+
+            if original_steps > self.config.num_train_timesteps:
+                raise ValueError(
+                    f"`original_steps`: {original_steps} cannot be larger than `self.config.train_timesteps`:"
+                    f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+                    f" maximal {self.config.num_train_timesteps} timesteps."
+                )
+
+            if num_inference_steps > original_steps:
+                raise ValueError(
+                    f"`num_inference_steps`: {num_inference_steps} cannot be larger than `original_inference_steps`:"
+                    f" {original_steps} because the final timestep schedule will be a subset of the"
+                    f" `original_inference_steps`-sized initial timestep schedule."
+                )
+
+            # LCM Timesteps Setting
+            # Currently, only linear spacing is supported.
+            c = self.config.num_train_timesteps // original_steps
+            # LCM Training Steps Schedule
+            lcm_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * c - 1
+            skipping_step = len(lcm_origin_timesteps) // num_inference_steps
+            # LCM Inference Steps Schedule
+            timesteps = lcm_origin_timesteps[::-skipping_step][:num_inference_steps]
+
+        self.timesteps = torch.from_numpy(timesteps.copy()).to(device=device, dtype=torch.long)
+
+        self._step_index = None
+
+    def get_scalings_for_boundary_condition_discrete(self, timestep):
+        self.sigma_data = 0.5  # Default: 0.5
+        scaled_timestep = timestep * self.config.timestep_scaling
+
+        c_skip = self.sigma_data**2 / (scaled_timestep**2 + self.sigma_data**2)
+        c_out = scaled_timestep / (scaled_timestep**2 + self.sigma_data**2) ** 0.5
+        return c_skip, c_out
+
+    def append_dims(self, x, target_dims):
+        """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
+        dims_to_append = target_dims - x.ndim
+        if dims_to_append < 0:
+            raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
+        return x[(...,) + (None,) * dims_to_append]
+
+    def extract_into_tensor(self, a, t, x_shape):
+        b, *_ = t.shape
+        out = a.gather(-1, t)
+        return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+    def step(
+        self,
+        model_output: torch.FloatTensor,
+        timestep: torch.Tensor,
+        sample: torch.FloatTensor,
+        generator: Optional[torch.Generator] = None,
+        return_dict: bool = True,
+    ) -> Union[LCMSingleStepSchedulerOutput, Tuple]:
+        """
+        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+        process from the learned model outputs (most often the predicted noise).
+
+        Args:
+            model_output (`torch.FloatTensor`):
+                The direct output from learned diffusion model.
+            timestep (`float`):
+                The current discrete timestep in the diffusion chain.
+            sample (`torch.FloatTensor`):
+                A current instance of a sample created by the diffusion process.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] or `tuple`.
+        Returns:
+            [`~schedulers.scheduling_utils.LCMSchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] is returned, otherwise a
+                tuple is returned where the first element is the sample tensor.
+        """
+        # 0. make sure everything is on the same device
+        alphas_cumprod = self.alphas_cumprod.to(sample.device)
+
+        # 1. compute alphas, betas
+        if timestep.ndim == 0:
+            timestep = timestep.unsqueeze(0)
+        alpha_prod_t = self.extract_into_tensor(alphas_cumprod, timestep, sample.shape)
+        beta_prod_t = 1 - alpha_prod_t
+
+        # 2. Get scalings for boundary conditions
+        c_skip, c_out = self.get_scalings_for_boundary_condition_discrete(timestep)
+        c_skip, c_out = [self.append_dims(x, sample.ndim) for x in [c_skip, c_out]]
+
+        # 3. Compute the predicted original sample x_0 based on the model parameterization
+        if self.config.prediction_type == "epsilon":  # noise-prediction
+            predicted_original_sample = (sample - torch.sqrt(beta_prod_t) * model_output) / torch.sqrt(alpha_prod_t)
+        elif self.config.prediction_type == "sample":  # x-prediction
+            predicted_original_sample = model_output
+        elif self.config.prediction_type == "v_prediction":  # v-prediction
+            predicted_original_sample = torch.sqrt(alpha_prod_t) * sample - torch.sqrt(beta_prod_t) * model_output
+        else:
+            raise ValueError(
+                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or"
+                " `v_prediction` for `LCMScheduler`."
+            )
+
+        # 4. Clip or threshold "predicted x_0"
+        if self.config.thresholding:
+            predicted_original_sample = self._threshold_sample(predicted_original_sample)
+        elif self.config.clip_sample:
+            predicted_original_sample = predicted_original_sample.clamp(
+                -self.config.clip_sample_range, self.config.clip_sample_range
+            )
+
+        # 5. Denoise model output using boundary conditions
+        denoised = c_out * predicted_original_sample + c_skip * sample
+
+        if not return_dict:
+            return (denoised, )
+
+        return LCMSingleStepSchedulerOutput(denoised=denoised)
+
+    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
+    def add_noise(
+        self,
+        original_samples: torch.FloatTensor,
+        noise: torch.FloatTensor,
+        timesteps: torch.IntTensor,
+    ) -> torch.FloatTensor:
+        # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+        alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype)
+        timesteps = timesteps.to(original_samples.device)
+
+        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+        while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+        while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+        noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+        return noisy_samples
+
+    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity
+    def get_velocity(
+        self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
+    ) -> torch.FloatTensor:
+        # Make sure alphas_cumprod and timestep have same device and dtype as sample
+        alphas_cumprod = self.alphas_cumprod.to(device=sample.device, dtype=sample.dtype)
+        timesteps = timesteps.to(sample.device)
+
+        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+        while len(sqrt_alpha_prod.shape) < len(sample.shape):
+            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+        while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+        velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+        return velocity
+
+    def __len__(self):
+        return self.config.num_train_timesteps
diff --git a/modules/instantir/sdxl_instantir.py b/modules/instantir/sdxl_instantir.py
new file mode 100644
index 000000000..b279d8445
--- /dev/null
+++ b/modules/instantir/sdxl_instantir.py
@@ -0,0 +1,1738 @@
+# Copyright 2024 The InstantX Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import (
+    CLIPImageProcessor,
+    CLIPTextModel,
+    CLIPTextModelWithProjection,
+    CLIPTokenizer,
+    CLIPVisionModelWithProjection,
+)
+
+from diffusers.utils.import_utils import is_invisible_watermark_available
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import (
+    FromSingleFileMixin,
+    IPAdapterMixin,
+    StableDiffusionXLLoraLoaderMixin,
+    TextualInversionLoaderMixin,
+)
+from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from diffusers.models.attention_processor import (
+    AttnProcessor2_0,
+    LoRAAttnProcessor2_0,
+    LoRAXFormersAttnProcessor,
+    XFormersAttnProcessor,
+)
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.schedulers import KarrasDiffusionSchedulers, LCMScheduler
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    deprecate,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+    convert_unet_state_dict_to_peft
+)
+from diffusers.utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+    from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
+
+from peft import LoraConfig, set_peft_model_state_dict
+from .aggregator import Aggregator
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> # !pip install diffusers pillow transformers accelerate
+        >>> import torch
+        >>> from PIL import Image
+
+        >>> from diffusers import DDPMScheduler
+        >>> from schedulers.lcm_single_step_scheduler import LCMSingleStepScheduler
+
+        >>> from module.ip_adapter.utils import load_adapter_to_pipe
+        >>> from pipelines.sdxl_instantir import InstantIRPipeline
+
+        >>> # download models under ./models
+        >>> dcp_adapter = f'./models/adapter.pt'
+        >>> previewer_lora_path = f'./models'
+        >>> instantir_path = f'./models/aggregator.pt'
+
+        >>> # load pretrained models
+        >>> pipe = InstantIRPipeline.from_pretrained(
+        ...     "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16
+        ... )
+        >>> # load adapter
+        >>> load_adapter_to_pipe(
+        ...     pipe,
+        ...     dcp_adapter,
+        ...     image_encoder_or_path = 'facebook/dinov2-large',
+        ... )
+        >>> # load previewer lora
+        >>> pipe.prepare_previewers(previewer_lora_path)
+        >>> pipe.scheduler = DDPMScheduler.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', subfolder="scheduler")
+        >>> lcm_scheduler = LCMSingleStepScheduler.from_config(pipe.scheduler.config)
+
+        >>> # load aggregator weights
+        >>> pretrained_state_dict = torch.load(instantir_path)
+        >>> pipe.aggregator.load_state_dict(pretrained_state_dict)
+
+        >>> # send to GPU and fp16
+        >>> pipe.to(device="cuda", dtype=torch.float16)
+        >>> pipe.aggregator.to(device="cuda", dtype=torch.float16)
+        >>> pipe.enable_model_cpu_offload()
+
+        >>> # load a broken image
+        >>> low_quality_image = Image.open('path/to/your-image').convert("RGB")
+
+        >>> # restoration
+        >>> image = pipe(
+        ...     image=low_quality_image,
+        ...     previewer_scheduler=lcm_scheduler,
+        ... ).images[0]
+        ```
+"""
+
+LCM_LORA_MODULES = [
+    "to_q",
+    "to_k",
+    "to_v",
+    "to_out.0",
+    "proj_in",
+    "proj_out",
+    "ff.net.0.proj",
+    "ff.net.2",
+    "conv1",
+    "conv2",
+    "conv_shortcut",
+    "downsamplers.0.conv",
+    "upsamplers.0.conv",
+    "time_emb_proj",
+]
+PREVIEWER_LORA_MODULES = [
+    "to_q",
+    "to_kv",
+    "0.to_out",
+    "attn1.to_k",
+    "attn1.to_v",
+    "to_k_ip",
+    "to_v_ip",
+    "ln_k_ip.linear",
+    "ln_v_ip.linear",
+    "to_out.0",
+    "proj_in",
+    "proj_out",
+    "ff.net.0.proj",
+    "ff.net.2",
+    "conv1",
+    "conv2",
+    "conv_shortcut",
+    "downsamplers.0.conv",
+    "upsamplers.0.conv",
+    "time_emb_proj",
+]
+
+
+def remove_attn2(model):
+    def recursive_find_module(name, module):
+        if not "up_blocks" in name and not "down_blocks" in name and not "mid_block" in name: return
+        elif "resnets" in name: return
+        if hasattr(module, "attn2"):
+            setattr(module, "attn2", None)
+            setattr(module, "norm2", None)
+            return
+        for sub_name, sub_module in module.named_children():
+            recursive_find_module(f"{name}.{sub_name}", sub_module)
+
+    for name, module in model.named_children():
+        recursive_find_module(name, module)
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+                timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+                must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class InstantIRPipeline(
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    StableDiffusionXLLoraLoaderMixin,
+    IPAdapterMixin,
+    FromSingleFileMixin,
+):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion XL with ControlNet guidance.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+        text_encoder ([`~transformers.CLIPTextModel`]):
+            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+        text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]):
+            Second frozen text-encoder
+            ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)).
+        tokenizer ([`~transformers.CLIPTokenizer`]):
+            A `CLIPTokenizer` to tokenize text.
+        tokenizer_2 ([`~transformers.CLIPTokenizer`]):
+            A `CLIPTokenizer` to tokenize text.
+        unet ([`UNet2DConditionModel`]):
+            A `UNet2DConditionModel` to denoise the encoded image latents.
+        controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
+            Provides additional conditioning to the `unet` during the denoising process. If you set multiple
+            ControlNets as a list, the outputs from each ControlNet are added together to create one combined
+            additional conditioning.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+            Whether the negative prompt embeddings should always be set to 0. Also see the config of
+            `stabilityai/stable-diffusion-xl-base-1-0`.
+        add_watermarker (`bool`, *optional*):
+            Whether to use the [invisible_watermark](https://github.com/ShieldMnt/invisible-watermark/) library to
+            watermark output images. If not defined, it defaults to `True` if the package is installed; otherwise no
+            watermarker is used.
+    """
+
+    # leave controlnet out on purpose because it iterates with unet
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+    _optional_components = [
+        "tokenizer",
+        "tokenizer_2",
+        "text_encoder",
+        "text_encoder_2",
+        "feature_extractor",
+        "image_encoder",
+    ]
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        text_encoder_2: CLIPTextModelWithProjection,
+        tokenizer: CLIPTokenizer,
+        tokenizer_2: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        aggregator: Aggregator = None,
+        force_zeros_for_empty_prompt: bool = True,
+        add_watermarker: Optional[bool] = None,
+        feature_extractor: CLIPImageProcessor = None,
+        image_encoder: CLIPVisionModelWithProjection = None,
+    ):
+        super().__init__()
+
+        if aggregator is None:
+            aggregator = Aggregator.from_unet(unet)
+        remove_attn2(aggregator)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            text_encoder_2=text_encoder_2,
+            tokenizer=tokenizer,
+            tokenizer_2=tokenizer_2,
+            unet=unet,
+            aggregator=aggregator,
+            scheduler=scheduler,
+            feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+        self.control_image_processor = VaeImageProcessor(
+            vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=True
+        )
+        add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+        if add_watermarker:
+            self.watermark = StableDiffusionXLWatermarker()
+        else:
+            self.watermark = None
+
+        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+
+    def prepare_previewers(self, previewer_lora_path: str, use_lcm=False):
+        if use_lcm:
+            lora_state_dict, alpha_dict = self.lora_state_dict(
+                previewer_lora_path,
+            )
+        else:
+            lora_state_dict, alpha_dict = self.lora_state_dict(
+                previewer_lora_path,
+                weight_name="previewer_lora_weights.bin"
+            )
+        unet_state_dict = {
+            f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")
+        }
+        unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
+        lora_state_dict = dict()
+        for k, v in unet_state_dict.items():
+            if "ip" in k:
+                k = k.replace("attn2", "attn2.processor")
+                lora_state_dict[k] = v
+            else:
+                lora_state_dict[k] = v
+        if alpha_dict:
+            lora_alpha = next(iter(alpha_dict.values()))
+        else:
+            lora_alpha = 1
+        logger.info(f"use lora alpha {lora_alpha}")
+        lora_config = LoraConfig(
+            r=64,
+            target_modules=LCM_LORA_MODULES if use_lcm else PREVIEWER_LORA_MODULES,
+            lora_alpha=lora_alpha,
+            lora_dropout=0.0,
+        )
+
+        adapter_name = "lcm" if use_lcm else "previewer"
+        self.unet.add_adapter(lora_config, adapter_name)
+        incompatible_keys = set_peft_model_state_dict(self.unet, lora_state_dict, adapter_name=adapter_name)
+        if incompatible_keys is not None:
+            # check only for unexpected keys
+            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+            missing_keys = getattr(incompatible_keys, "missing_keys", None)
+            if unexpected_keys:
+                raise ValueError(
+                    f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                    f" {unexpected_keys}. "
+                )
+        self.unet.disable_adapters()
+
+        return lora_alpha
+    
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: str,
+        prompt_2: Optional[str] = None,
+        device: Optional[torch.device] = None,
+        num_images_per_prompt: int = 1,
+        do_classifier_free_guidance: bool = True,
+        negative_prompt: Optional[str] = None,
+        negative_prompt_2: Optional[str] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+        clip_skip: Optional[int] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+                less than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            lora_scale (`float`, *optional*):
+                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        device = device or self._execution_device
+
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder, lora_scale)
+
+            if self.text_encoder_2 is not None:
+                if not USE_PEFT_BACKEND:
+                    adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+                else:
+                    scale_lora_layers(self.text_encoder_2, lora_scale)
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        if prompt is not None:
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # Define tokenizers and text encoders
+        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+        text_encoders = (
+            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+        )
+
+        if prompt_embeds is None:
+            prompt_2 = prompt_2 or prompt
+            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+            # textual inversion: process multi-vector tokens if necessary
+            prompt_embeds_list = []
+            prompts = [prompt, prompt_2]
+            for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+                text_inputs = tokenizer(
+                    prompt,
+                    padding="max_length",
+                    max_length=tokenizer.model_max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                text_input_ids = text_inputs.input_ids
+                untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                    text_input_ids, untruncated_ids
+                ):
+                    removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+                    logger.warning(
+                        "The following part of your input was truncated because CLIP can only handle sequences up to"
+                        f" {tokenizer.model_max_length} tokens: {removed_text}"
+                    )
+
+                prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                pooled_prompt_embeds = prompt_embeds[0]
+                if clip_skip is None:
+                    prompt_embeds = prompt_embeds.hidden_states[-2]
+                else:
+                    # "2" because SDXL always indexes from the penultimate layer.
+                    prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+                prompt_embeds_list.append(prompt_embeds)
+
+            prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+        # get unconditional embeddings for classifier free guidance
+        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+        if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+            negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+            negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+        elif do_classifier_free_guidance and negative_prompt_embeds is None:
+            negative_prompt = negative_prompt or ""
+            negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+            # normalize str to list
+            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+            negative_prompt_2 = (
+                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+            )
+
+            uncond_tokens: List[str]
+            if prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = [negative_prompt, negative_prompt_2]
+
+            negative_prompt_embeds_list = []
+            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+                if isinstance(self, TextualInversionLoaderMixin):
+                    negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+                max_length = prompt_embeds.shape[1]
+                uncond_input = tokenizer(
+                    negative_prompt,
+                    padding="max_length",
+                    max_length=max_length,
+                    truncation=True,
+                    return_tensors="pt",
+                )
+
+                negative_prompt_embeds = text_encoder(
+                    uncond_input.input_ids.to(device),
+                    output_hidden_states=True,
+                )
+                # We are only ALWAYS interested in the pooled output of the final text encoder
+                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+                negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+                negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+            negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+        if self.text_encoder_2 is not None:
+            prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+        else:
+            prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            if self.text_encoder_2 is not None:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+            else:
+                negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+            bs_embed * num_images_per_prompt, -1
+        )
+        if do_classifier_free_guidance:
+            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+                bs_embed * num_images_per_prompt, -1
+            )
+
+        if self.text_encoder is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
+        if self.text_encoder_2 is not None:
+            if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            if isinstance(self.image_encoder, CLIPVisionModelWithProjection):
+                # CLIP image encoder.
+                image_embeds = self.image_encoder(image).image_embeds
+                image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+                uncond_image_embeds = torch.zeros_like(image_embeds)
+            else:
+                # DINO image encoder.
+                image_embeds = self.image_encoder(image).last_hidden_state
+                image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+                uncond_image_embeds = self.image_encoder(
+                    torch.zeros_like(image)
+                ).last_hidden_state
+                uncond_image_embeds = uncond_image_embeds.repeat_interleave(
+                    num_images_per_prompt, dim=0
+                )
+
+            return image_embeds, uncond_image_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+    def prepare_ip_adapter_image_embeds(
+        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+    ):
+        if ip_adapter_image_embeds is None:
+            if not isinstance(ip_adapter_image, list):
+                ip_adapter_image = [ip_adapter_image]
+
+            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+                if isinstance(ip_adapter_image[0], list):
+                    raise ValueError(
+                        f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+                    )
+                else:
+                    logger.warning(
+                        f"Got {len(ip_adapter_image)} images for {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+                        " By default, these images will be sent to each IP-Adapter. If this is not your use-case, please specify `ip_adapter_image` as a list of image-list, with"
+                        f" length equals to the number of IP-Adapters."
+                    )
+                    ip_adapter_image = [ip_adapter_image] * len(self.unet.encoder_hid_proj.image_projection_layers)
+
+            image_embeds = []
+            for single_ip_adapter_image, image_proj_layer in zip(
+                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+            ):
+                output_hidden_state = isinstance(self.image_encoder, CLIPVisionModelWithProjection) and not isinstance(image_proj_layer, ImageProjection)
+                single_image_embeds, single_negative_image_embeds = self.encode_image(
+                    single_ip_adapter_image, device, 1, output_hidden_state
+                )
+                single_image_embeds = torch.stack([single_image_embeds] * (num_images_per_prompt//single_image_embeds.shape[0]), dim=0)
+                single_negative_image_embeds = torch.stack(
+                    [single_negative_image_embeds] * (num_images_per_prompt//single_negative_image_embeds.shape[0]), dim=0
+                )
+
+                if do_classifier_free_guidance:
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                    single_image_embeds = single_image_embeds.to(device)
+
+                image_embeds.append(single_image_embeds)
+        else:
+            repeat_dims = [1]
+            image_embeds = []
+            for single_image_embeds in ip_adapter_image_embeds:
+                if do_classifier_free_guidance:
+                    single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                    single_negative_image_embeds = single_negative_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+                    )
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                else:
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                image_embeds.append(single_image_embeds)
+
+        return image_embeds
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        image,
+        callback_steps,
+        negative_prompt=None,
+        negative_prompt_2=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        pooled_prompt_embeds=None,
+        ip_adapter_image=None,
+        ip_adapter_image_embeds=None,
+        negative_pooled_prompt_embeds=None,
+        controlnet_conditioning_scale=1.0,
+        control_guidance_start=0.0,
+        control_guidance_end=1.0,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+        if negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if prompt_embeds is not None and pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+            )
+
+        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+            raise ValueError(
+                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+            )
+
+        # Check `image`
+        is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+            self.aggregator, torch._dynamo.eval_frame.OptimizedModule
+        )
+        if (
+            isinstance(self.aggregator, Aggregator)
+            or is_compiled
+            and isinstance(self.aggregator._orig_mod, Aggregator)
+        ):
+            self.check_image(image, prompt, prompt_embeds)
+        else:
+            assert False
+
+        if control_guidance_start >= control_guidance_end:
+            raise ValueError(
+                f"control guidance start: {control_guidance_start} cannot be larger or equal to control guidance end: {control_guidance_end}."
+            )
+        if control_guidance_start < 0.0:
+            raise ValueError(f"control guidance start: {control_guidance_start} can't be smaller than 0.")
+        if control_guidance_end > 1.0:
+            raise ValueError(f"control guidance end: {control_guidance_end} can't be larger than 1.0.")
+
+        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+            raise ValueError(
+                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+            )
+
+        if ip_adapter_image_embeds is not None:
+            if not isinstance(ip_adapter_image_embeds, list):
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+                )
+            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+                )
+
+    # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
+    def check_image(self, image, prompt, prompt_embeds):
+        image_is_pil = isinstance(image, PIL.Image.Image)
+        image_is_tensor = isinstance(image, torch.Tensor)
+        image_is_np = isinstance(image, np.ndarray)
+        image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+        image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+        image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+        if (
+            not image_is_pil
+            and not image_is_tensor
+            and not image_is_np
+            and not image_is_pil_list
+            and not image_is_tensor_list
+            and not image_is_np_list
+        ):
+            raise TypeError(
+                f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+            )
+
+        if image_is_pil:
+            image_batch_size = 1
+        else:
+            image_batch_size = len(image)
+
+        if prompt is not None and isinstance(prompt, str):
+            prompt_batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            prompt_batch_size = len(prompt)
+        elif prompt_embeds is not None:
+            prompt_batch_size = prompt_embeds.shape[0]
+
+        if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+            raise ValueError(
+                f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+            )
+
+    # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image
+    def prepare_image(
+        self,
+        image,
+        width,
+        height,
+        batch_size,
+        num_images_per_prompt,
+        device,
+        dtype,
+        do_classifier_free_guidance=False,
+    ):
+        image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+        image_batch_size = image.shape[0]
+
+        if image_batch_size == 1:
+            repeat_by = batch_size
+        else:
+            # image batch size is the same as prompt batch size
+            repeat_by = num_images_per_prompt
+
+        image = image.repeat_interleave(repeat_by, dim=0)
+
+        image = image.to(device=device, dtype=dtype)
+
+        return image
+
+    @torch.no_grad()
+    def init_latents(self, latents, generator, timestep):
+        noise = torch.randn(latents.shape, generator=generator[0] if isinstance(generator, list) else generator, device=self.vae.device, dtype=self.vae.dtype, layout=torch.strided)
+        bsz = latents.shape[0]
+        timestep = torch.tensor([timestep]*bsz, device=self.vae.device)
+        # Note that the latents will be scaled aleady by scheduler.add_noise
+        latents = self.scheduler.add_noise(latents, noise, timestep)
+        return latents
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (
+            batch_size,
+            num_channels_latents,
+            int(height) // self.vae_scale_factor,
+            int(width) // self.vae_scale_factor,
+        )
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
+    def _get_add_time_ids(
+        self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
+    ):
+        add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+        passed_add_embed_dim = (
+            self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+        )
+        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+        if expected_add_embed_dim != passed_add_embed_dim:
+            raise ValueError(
+                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+            )
+
+        add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+        return add_time_ids
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+    def upcast_vae(self):
+        dtype = self.vae.dtype
+        self.vae.to(dtype=torch.float32)
+        use_torch_2_0_or_xformers = isinstance(
+            self.vae.decoder.mid_block.attentions[0].processor,
+            (
+                AttnProcessor2_0,
+                XFormersAttnProcessor,
+                LoRAXFormersAttnProcessor,
+                LoRAAttnProcessor2_0,
+            ),
+        )
+        # if xformers or torch_2_0 is used attention block does not need
+        # to be in float32 which can save lots of memory
+        if use_torch_2_0_or_xformers:
+            self.vae.post_quant_conv.to(dtype)
+            self.vae.decoder.conv_in.to(dtype)
+            self.vae.decoder.mid_block.to(dtype)
+
+    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(
+        self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32
+    ) -> torch.FloatTensor:
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            w (`torch.Tensor`):
+                Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
+            embedding_dim (`int`, *optional*, defaults to 512):
+                Dimension of the embeddings to generate.
+            dtype (`torch.dtype`, *optional*, defaults to `torch.float32`):
+                Data type of the generated embeddings.
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def guidance_rescale(self):
+        return self._guidance_rescale
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+    # corresponds to doing no classifier free guidance.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+    @property
+    def cross_attention_kwargs(self):
+        return self._cross_attention_kwargs
+
+    @property
+    def denoising_end(self):
+        return self._denoising_end
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        image: PipelineImageInput = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 30,
+        timesteps: List[int] = None,
+        denoising_end: Optional[float] = None,
+        guidance_scale: float = 7.0,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        negative_prompt_2: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        save_preview_row: bool = False,
+        init_latents_with_lq: bool = True,
+        multistep_restore: bool = False,
+        adastep_restore: bool = False,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        controlnet_conditioning_scale: float = 1.0,
+        control_guidance_start: float = 0.0,
+        control_guidance_end: float = 1.0,
+        preview_start: float = 0.0,
+        preview_end: float = 1.0,
+        original_size: Tuple[int, int] = None,
+        crops_coords_top_left: Tuple[int, int] = (0, 0),
+        target_size: Tuple[int, int] = None,
+        negative_original_size: Optional[Tuple[int, int]] = None,
+        negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+        negative_target_size: Optional[Tuple[int, int]] = None,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        previewer_scheduler: KarrasDiffusionSchedulers = None,
+        reference_latents: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                used in both text-encoders.
+            image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+                    `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+                The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+                specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
+                accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+                and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
+                `init`, images must be passed as a list such that each element of the list can be correctly batched for
+                input to a single ControlNet.
+            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The height in pixels of the generated image. Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+                The width in pixels of the generated image. Anything below 512 pixels won't work well for
+                [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+                and checkpoints that are not specifically fine-tuned on low resolutions.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            denoising_end (`float`, *optional*):
+                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+                completed before it is intentionally prematurely terminated. As a result, the returned sample will
+                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+                Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+            guidance_scale (`float`, *optional*, defaults to 5.0):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. This is sent to `tokenizer_2`
+                and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, pooled text embeddings are generated from `prompt` input argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs (prompt
+                weighting). If not provided, pooled `negative_prompt_embeds` are generated from `negative_prompt` input
+                argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
+                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
+                contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
+                provided, embeddings are computed from the `ip_adapter_image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+                to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+                the corresponding scale as a list.
+            control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+                The percentage of total steps at which the ControlNet starts applying.
+            control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+                The percentage of total steps at which the ControlNet stops applying.
+            original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+                `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+                explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                For most cases, `target_size` should be set to the desired height and width of the generated image. If
+                not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+            negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+                To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+                micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+                To negatively condition the generation process based on a target image resolution. It should be as same
+                as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+                information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned containing the output images.
+        """
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+            )
+
+        aggregator = self.aggregator._orig_mod if is_compiled_module(self.aggregator) else self.aggregator
+        if not isinstance(ip_adapter_image, list):
+            ip_adapter_image = [ip_adapter_image] if ip_adapter_image is not None else [image]
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            image,
+            callback_steps,
+            negative_prompt,
+            negative_prompt_2,
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            negative_pooled_prompt_embeds,
+            controlnet_conditioning_scale,
+            control_guidance_start,
+            control_guidance_end,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._denoising_end = denoising_end
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            if not isinstance(image, PIL.Image.Image):
+                batch_size = len(image)
+            else:
+                batch_size = 1
+            prompt = [prompt] * batch_size
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+            assert batch_size == len(image) or (isinstance(image, PIL.Image.Image) or len(image) == 1)
+        else:
+            batch_size = prompt_embeds.shape[0]
+            assert batch_size == len(image) or (isinstance(image, PIL.Image.Image) or len(image) == 1)
+
+        device = self._execution_device
+
+        # 3.1 Encode input prompt
+        text_encoder_lora_scale = (
+            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            negative_prompt_embeds,
+            pooled_prompt_embeds,
+            negative_pooled_prompt_embeds,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            negative_prompt_2=negative_prompt_2,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+            clip_skip=self.clip_skip,
+        )
+        # 3.2 Encode ip_adapter_image
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+            image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+        # 4. Prepare image
+        image = self.prepare_image(
+            image=image,
+            width=width,
+            height=height,
+            batch_size=batch_size * num_images_per_prompt,
+            num_images_per_prompt=num_images_per_prompt,
+            device=device,
+            dtype=aggregator.dtype,
+            do_classifier_free_guidance=self.do_classifier_free_guidance,
+        )
+        height, width = image.shape[-2:]
+        if image.shape[1] != 4:
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+            if needs_upcasting:
+                image = image.float()
+                self.vae.to(dtype=torch.float32)
+            image = self.vae.encode(image).latent_dist.sample()
+            image = image * self.vae.config.scaling_factor
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+                image = image.to(dtype=torch.float16)
+        else:
+            height = int(height * self.vae_scale_factor)
+            width = int(width * self.vae_scale_factor)
+
+        # 5. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+        # 6. Prepare latent variables
+        if init_latents_with_lq:
+            latents = self.init_latents(image, generator, timesteps[0])
+        else:
+            num_channels_latents = self.unet.config.in_channels
+            latents = self.prepare_latents(
+                batch_size * num_images_per_prompt,
+                num_channels_latents,
+                height,
+                width,
+                prompt_embeds.dtype,
+                device,
+                generator,
+                latents,
+            )
+
+        # 6.5 Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        # 7.1 Create tensor stating which controlnets to keep
+        controlnet_keep = []
+        previewing = []
+        for i in range(len(timesteps)):
+            keeps = 1.0 - float(i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end)
+            controlnet_keep.append(keeps)
+            use_preview = 1.0 - float(i / len(timesteps) < preview_start or (i + 1) / len(timesteps) > preview_end)
+            previewing.append(use_preview)
+        if isinstance(controlnet_conditioning_scale, list):
+            assert len(controlnet_conditioning_scale) == len(timesteps), f"{len(controlnet_conditioning_scale)} controlnet scales do not match number of sampling steps {len(timesteps)}"
+        else:
+            controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet_keep)
+
+        # 7.2 Prepare added time ids & embeddings
+        original_size = original_size or (height, width)
+        target_size = target_size or (height, width)
+
+        add_text_embeds = pooled_prompt_embeds
+        if self.text_encoder_2 is None:
+            text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+        else:
+            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+        add_time_ids = self._get_add_time_ids(
+            original_size,
+            crops_coords_top_left,
+            target_size,
+            dtype=prompt_embeds.dtype,
+            text_encoder_projection_dim=text_encoder_projection_dim,
+        )
+
+        if negative_original_size is not None and negative_target_size is not None:
+            negative_add_time_ids = self._get_add_time_ids(
+                negative_original_size,
+                negative_crops_coords_top_left,
+                negative_target_size,
+                dtype=prompt_embeds.dtype,
+                text_encoder_projection_dim=text_encoder_projection_dim,
+            )
+        else:
+            negative_add_time_ids = add_time_ids
+
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+            add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+            add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+            image = torch.cat([image] * 2, dim=0)
+
+        prompt_embeds = prompt_embeds.to(device)
+        add_text_embeds = add_text_embeds.to(device)
+        add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+        # 8. Denoising loop
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+        # 8.1 Apply denoising_end
+        if (
+            self.denoising_end is not None
+            and isinstance(self.denoising_end, float)
+            and self.denoising_end > 0
+            and self.denoising_end < 1
+        ):
+            discrete_timestep_cutoff = int(
+                round(
+                    self.scheduler.config.num_train_timesteps
+                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+                )
+            )
+            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+            timesteps = timesteps[:num_inference_steps]
+
+        is_unet_compiled = is_compiled_module(self.unet)
+        is_aggregator_compiled = is_compiled_module(self.aggregator)
+        is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
+        previewer_mean = torch.zeros_like(latents)
+        unet_mean = torch.zeros_like(latents)
+        preview_factor = torch.ones(
+            (latents.shape[0], *((1,) * (len(latents.shape) - 1))), dtype=latents.dtype, device=latents.device
+        )
+
+        self._num_timesteps = len(timesteps)
+        preview_row = []
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                # Relevant thread:
+                # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+                if (is_unet_compiled and is_aggregator_compiled) and is_torch_higher_equal_2_1:
+                    torch._inductor.cudagraph_mark_step_begin()
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+                prev_t = t
+                unet_model_input = latent_model_input
+
+                added_cond_kwargs = {
+                    "text_embeds": add_text_embeds,
+                    "time_ids": add_time_ids,
+                    "image_embeds": image_embeds
+                }
+                aggregator_added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+
+                # prepare time_embeds in advance as adapter input
+                cross_attention_t_emb = self.unet.get_time_embed(sample=latent_model_input, timestep=t)
+                cross_attention_emb = self.unet.time_embedding(cross_attention_t_emb, timestep_cond)
+                cross_attention_aug_emb = None
+
+                cross_attention_aug_emb = self.unet.get_aug_embed(
+                    emb=cross_attention_emb,
+                    encoder_hidden_states=prompt_embeds,
+                    added_cond_kwargs=added_cond_kwargs
+                )
+
+                cross_attention_emb = cross_attention_emb + cross_attention_aug_emb if cross_attention_aug_emb is not None else cross_attention_emb
+
+                if self.unet.time_embed_act is not None:
+                    cross_attention_emb = self.unet.time_embed_act(cross_attention_emb)
+
+                current_cross_attention_kwargs = {"temb": cross_attention_emb}
+                if cross_attention_kwargs is not None:
+                    for k,v in cross_attention_kwargs.items():
+                        current_cross_attention_kwargs[k] = v
+                self._cross_attention_kwargs = current_cross_attention_kwargs
+
+                # adaptive restoration factors
+                adaRes_scale = preview_factor.to(latent_model_input.dtype).clamp(0.0, controlnet_conditioning_scale[i])
+                cond_scale = adaRes_scale * controlnet_keep[i]
+                cond_scale = torch.cat([cond_scale] * 2) if self.do_classifier_free_guidance else cond_scale
+
+                if (cond_scale>0.1).sum().item() > 0:
+                    if previewing[i] > 0:
+                        # preview with LCM
+                        self.unet.enable_adapters()
+                        preview_noise = self.unet(
+                            latent_model_input,
+                            t,
+                            encoder_hidden_states=prompt_embeds,
+                            timestep_cond=timestep_cond,
+                            cross_attention_kwargs=self.cross_attention_kwargs,
+                            added_cond_kwargs=added_cond_kwargs,
+                            return_dict=False,
+                        )[0]
+                        preview_latent = previewer_scheduler.step(
+                            preview_noise,
+                            t.to(dtype=torch.int64),
+                            # torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents,
+                            latent_model_input,     # scaled latents here for compatibility
+                            return_dict=False
+                        )[0]
+                        self.unet.disable_adapters()
+
+                        if self.do_classifier_free_guidance:
+                            preview_row.append(preview_latent.chunk(2)[1].to('cpu'))
+                        else:
+                            preview_row.append(preview_latent.to('cpu'))
+                        # Prepare 2nd order step.
+                        if multistep_restore and i+1 < len(timesteps):
+                            noise_preview = preview_noise.chunk(2)[1] if self.do_classifier_free_guidance else preview_noise
+                            first_step = self.scheduler.step(
+                                noise_preview, t, latents,
+                                **extra_step_kwargs, return_dict=True, step_forward=False
+                            )
+                            prev_t = timesteps[i + 1]
+                            unet_model_input = torch.cat([first_step.prev_sample] * 2) if self.do_classifier_free_guidance else first_step.prev_sample
+                            unet_model_input = self.scheduler.scale_model_input(unet_model_input, prev_t, heun_step=True)
+
+                    elif reference_latents is not None:
+                        preview_latent = torch.cat([reference_latents] * 2) if self.do_classifier_free_guidance else reference_latents
+                    else:
+                        preview_latent = image
+
+                    # Add fresh noise
+                    # preview_noise = torch.randn_like(preview_latent)
+                    # preview_latent = self.scheduler.add_noise(preview_latent, preview_noise, t)
+
+                    preview_latent=preview_latent.to(dtype=next(aggregator.parameters()).dtype)
+
+                    # Aggregator inference
+                    down_block_res_samples, mid_block_res_sample = aggregator(
+                        image,
+                        prev_t,
+                        encoder_hidden_states=prompt_embeds,
+                        controlnet_cond=preview_latent,
+                        # conditioning_scale=cond_scale,
+                        added_cond_kwargs=aggregator_added_cond_kwargs,
+                        return_dict=False,
+                    )
+
+                # aggregator features scaling
+                down_block_res_samples = [sample*cond_scale for sample in down_block_res_samples]
+                mid_block_res_sample = mid_block_res_sample*cond_scale
+
+                # predict the noise residual
+                noise_pred = self.unet(
+                    unet_model_input,
+                    prev_t,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
+                    down_block_additional_residuals=down_block_res_samples,
+                    mid_block_additional_residual=mid_block_res_sample,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                unet_step = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=True)
+                latents = unet_step.prev_sample
+
+                # Update adaRes factors
+                unet_pred_latent = unet_step.pred_original_sample
+
+                # Adaptive restoration.
+                if adastep_restore:
+                    pred_x0_l2 = ((preview_latent[latents.shape[0]:].float()-unet_pred_latent.float())).pow(2).sum(dim=(1,2,3))
+                    previewer_l2 = ((preview_latent[latents.shape[0]:].float()-previewer_mean.float())).pow(2).sum(dim=(1,2,3))
+                    # unet_l2 = ((unet_pred_latent.float()-unet_mean.float())).pow(2).sum(dim=(1,2,3)).sqrt()
+                    # l2_error = (((preview_latent[latents.shape[0]:]-previewer_mean) - (unet_pred_latent-unet_mean))).pow(2).mean(dim=(1,2,3))
+                    # preview_error = torch.nn.functional.cosine_similarity(preview_latent[latents.shape[0]:].reshape(latents.shape[0], -1), unet_pred_latent.reshape(latents.shape[0],-1))
+                    previewer_mean = preview_latent[latents.shape[0]:]
+                    unet_mean = unet_pred_latent
+                    preview_factor = (pred_x0_l2 / previewer_l2).reshape(-1, 1, 1, 1)
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if not output_type == "latent":
+            # make sure the VAE is in float32 mode, as it overflows in float16
+            needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+            if needs_upcasting:
+                self.upcast_vae()
+                latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+            # unscale/denormalize the latents
+            # denormalize with the mean and std if available and not None
+            has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+            has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+            if has_latents_mean and has_latents_std:
+                latents_mean = (
+                    torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents_std = (
+                    torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+                )
+                latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+            else:
+                latents = latents / self.vae.config.scaling_factor
+
+            image = self.vae.decode(latents, return_dict=False)[0]
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+        else:
+            image = latents
+
+        if not output_type == "latent":
+            # apply watermark if available
+            if self.watermark is not None:
+                image = self.watermark.apply_watermark(image)
+
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        if save_preview_row:
+            preview_image_row = []
+            if needs_upcasting:
+                self.upcast_vae()
+            for preview_latents in preview_row:
+                preview_latents = preview_latents.to(device=self.device, dtype=next(iter(self.vae.post_quant_conv.parameters())).dtype)
+                if has_latents_mean and has_latents_std:
+                    latents_mean = (
+                        torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(preview_latents.device, preview_latents.dtype)
+                    )
+                    latents_std = (
+                        torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(preview_latents.device, preview_latents.dtype)
+                    )
+                    preview_latents = preview_latents * latents_std / self.vae.config.scaling_factor + latents_mean
+                else:
+                    preview_latents = preview_latents / self.vae.config.scaling_factor
+
+                preview_image = self.vae.decode(preview_latents, return_dict=False)[0]
+                preview_image = self.image_processor.postprocess(preview_image, output_type=output_type)
+                preview_image_row.append(preview_image)
+
+            # cast back to fp16 if needed
+            if needs_upcasting:
+                self.vae.to(dtype=torch.float16)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            if save_preview_row:
+                return (image, preview_image_row)
+            return (image,)
+
+        return StableDiffusionXLPipelineOutput(images=image)
diff --git a/modules/sd_models.py b/modules/sd_models.py
index 1a17660a3..354ceea7d 100644
--- a/modules/sd_models.py
+++ b/modules/sd_models.py
@@ -1055,6 +1055,7 @@ def set_diffuser_pipe(pipe, new_pipe_type):
         'OmniGenPipeline',
         'StableDiffusion3ControlNetPipeline',
         'StableDiffusionXLPuLIDPipeline',
+        'InstantIRPipeline',
     ]
 
     n = getattr(pipe.__class__, '__name__', '')
diff --git a/scripts/instantir.py b/scripts/instantir.py
new file mode 100644
index 000000000..4c7ce77b7
--- /dev/null
+++ b/scripts/instantir.py
@@ -0,0 +1,97 @@
+import gradio as gr
+import torch
+import diffusers
+from huggingface_hub import hf_hub_download
+from modules import scripts, processing, shared, sd_models, devices, ipadapter
+
+
+class Script(scripts.Script):
+    def __init__(self):
+        super().__init__()
+        self.orig_pipe = None
+        self.orig_ip_unapply = None
+
+    def title(self):
+        return 'InstantIR'
+
+    def show(self, is_img2img):
+        return is_img2img if shared.native else False
+
+    def ui(self, _is_img2img): # ui elements
+        with gr.Row():
+            gr.HTML('<a href="https://github.com/instantX-research/InstantIR">&nbsp InstantIR: Image Restoration</a><br>')
+        with gr.Row():
+            start = gr.Slider(label='Preview start', minimum=0.0, maximum=1.0, step=0.01, value=0.0)
+            end = gr.Slider(label='Preview end', minimum=0.0, maximum=1.0, step=0.01, value=1.0)
+        with gr.Row():
+            hq = gr.Checkbox(label='HQ init latents', value=False)
+            multistep = gr.Checkbox(label='Multistep restore', value=False)
+            adastep = gr.Checkbox(label='Adaptive restore', value=False)
+        with gr.Row():
+            image = gr.Image(label='Override guidance image')
+        return [start, end, hq, multistep, adastep, image]
+
+    def run(self, p: processing.StableDiffusionProcessing, *args): # pylint: disable=arguments-differ
+        supported_model_list = ['sdxl']
+        if not hasattr(p, 'init_images') or len(p.init_images) == 0:
+            shared.log.warning('InstantIR: no image')
+            return None
+        if shared.sd_model_type not in supported_model_list and shared.sd_model.__class__.__name__ != "InstantIRPipeline":
+            shared.log.warning(f'InstantIR: class={shared.sd_model.__class__.__name__} model={shared.sd_model_type} required={supported_model_list}')
+            return None
+        start, end, hq, multistep, adastep, image = args
+        from modules import instantir as ir
+        if shared.sd_model_type == "sdxl":
+            if shared.sd_model.__class__.__name__ != "InstantIRPipeline":
+                self.orig_pipe = shared.sd_model
+                self.orig_ip_unapply = ipadapter.unapply
+                shared.sd_model = sd_models.switch_pipe(ir.InstantIRPipeline, shared.sd_model)
+                adapter_file = hf_hub_download('InstantX/InstantIR', subfolder='models', filename='adapter.pt', cache_dir=shared.opts.hfcache_dir)
+                aggregator_file = hf_hub_download('InstantX/InstantIR', subfolder='models', filename='aggregator.pt', cache_dir=shared.opts.hfcache_dir)
+                previewer_file = hf_hub_download('InstantX/InstantIR', subfolder='models', filename='previewer_lora_weights.bin', cache_dir=shared.opts.hfcache_dir)
+                shared.log.debug(f'InstantIR: adapter="{adapter_file}" aggregator="{aggregator_file}" previewer="{previewer_file}"')
+                ir.load_adapter_to_pipe(
+                    pipe=shared.sd_model,
+                    pretrained_model_path_or_dict=adapter_file,
+                    image_encoder_or_path='facebook/dinov2-large',
+                    use_lcm=False,
+                    use_adaln=True,
+                    )
+                shared.sd_model.prepare_previewers(previewer_file)
+                shared.sd_model.scheduler = diffusers.DDPMScheduler.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', subfolder="scheduler")
+                pretrained_state_dict = torch.load(aggregator_file)
+                shared.sd_model.aggregator.load_state_dict(pretrained_state_dict)
+                shared.sd_model.aggregator.to(device=devices.device, dtype=devices.dtype)
+
+        shared.log.info(f'InstantIR: class={shared.sd_model.__class__.__name__} start={start} end={end} multistep={multistep} adastep={adastep} hq={hq} cache={shared.opts.hfcache_dir}')
+        p.sampler_name = 'Default' # ir has its own sampler
+        p.init() # run init early to take care of resizing
+        p.task_args['previewer_scheduler'] = ir.LCMSingleStepScheduler.from_config(shared.sd_model.scheduler.config)
+        p.task_args['image'] = p.init_images
+        p.task_args['save_preview_row'] = False
+        p.task_args['init_latents_with_lq'] = not hq
+        p.task_args['multistep_restore'] = multistep
+        p.task_args['adastep_restore'] = adastep
+        p.task_args['preview_start'] = start
+        p.task_args['preview_end'] = end
+        p.task_args['ip_adapter_image'] = image
+        p.extra_generation_params["InstantIR"] = f'Start={start} End={end} HQ={hq} Multistep={multistep} Adastep={adastep}'
+        ipadapter.unapply = lambda x: x # disable as main processing unloads ipadapter as it thinks its not needed
+        devices.torch_gc()
+
+    def after(self, p: processing.StableDiffusionProcessing, processed: processing.Processed, *args): # pylint: disable=arguments-differ, unused-argument
+        # TODO instantir is a mess to unload
+        """
+        if self.orig_pipe is None:
+            return processed
+        if hasattr(shared.sd_model, 'aggregator'):
+            shared.sd_model.aggregator = None
+        shared.log.debug(f'InstantIR restore: class={shared.sd_model.__class__.__name__}')
+        shared.sd_model = self.orig_pipe
+        self.orig_pipe = None
+        shared.sd_model.unet.register_to_config(encoder_hid_dim_type=None)
+        ipadapter.unapply = self.orig_ip_unapply
+        ipadapter.unapply(shared.sd_model)
+        devices.torch_gc()
+        """
+        return processed