From ccb1ded7d0c201c8e122410c96a4ba564abcd3be Mon Sep 17 00:00:00 2001 From: Vladimir Mandic Date: Sat, 15 Nov 2025 11:56:56 -0500 Subject: [PATCH] integrate kanvas with server-side masking Signed-off-by: Vladimir Mandic --- CHANGELOG.md | 2 +- extensions-builtin/sdnext-kanvas | 2 +- extensions-builtin/sdnext-modernui | 2 +- modules/ben2/__init__.py | 24 + modules/ben2/ben2_model.py | 1290 ++++++++++++++++++++++++++++ modules/masking.py | 223 ++--- modules/processing_class.py | 45 +- modules/ui_control.py | 7 +- 8 files changed, 1469 insertions(+), 126 deletions(-) create mode 100644 modules/ben2/__init__.py create mode 100644 modules/ben2/ben2_model.py diff --git a/CHANGELOG.md b/CHANGELOG.md index 779e107e6..21d9bfdb1 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -35,7 +35,7 @@ And a first cloud model with **Google's Nano Banana** - **auth**: strong-enforce auth check on all api endpoints - **amdgpu**: prefer rocm-on-windows over zluda - **Internal** - - torch: update to `torch==2.9.1` for cuda and ipex backends + - torch: update to `torch==2.9.1` for cuda, ipex, openvino, rocm backends - attention: refactor attention handling - scripts: remove obsolete video scripts - lint: update global lint rules diff --git a/extensions-builtin/sdnext-kanvas b/extensions-builtin/sdnext-kanvas index ec77ebd4a..ce162ca6c 160000 --- a/extensions-builtin/sdnext-kanvas +++ b/extensions-builtin/sdnext-kanvas @@ -1 +1 @@ -Subproject commit ec77ebd4a50ad5bdc977a7f163c623c9f7bc22e1 +Subproject commit ce162ca6cdb434e8cc4cdf9283a7b3852661e345 diff --git a/extensions-builtin/sdnext-modernui b/extensions-builtin/sdnext-modernui index 404a1e53a..e61bc4ee5 160000 --- a/extensions-builtin/sdnext-modernui +++ b/extensions-builtin/sdnext-modernui @@ -1 +1 @@ -Subproject commit 404a1e53a85d64a64c499690c71360735b710a8d +Subproject commit e61bc4ee5b3233f04e533a0c883cd6adbb5cedb2 diff --git a/modules/ben2/__init__.py b/modules/ben2/__init__.py new file mode 100644 index 000000000..f9375a990 --- /dev/null +++ b/modules/ben2/__init__.py @@ -0,0 +1,24 @@ +model = None + + +def remove(image, refine: bool = True): + global model # pylint: disable=global-statement + from modules import shared, devices + + if model is None: + from huggingface_hub import hf_hub_download + from .ben2_model import BEN_Base + model = BEN_Base() + model_file = hf_hub_download( + repo_id='PramaLLC/BEN2', + filename='BEN2_Base.pth', + cache_dir=shared.opts.hfcache_dir) + model.loadcheckpoints(model_file) + model = model.to(device=devices.device, dtype=devices.dtype).eval() + + model = model.to(device=devices.device) + foreground = model.inference(image, refine_foreground=refine) + model = model.to(device=devices.cpu) + if foreground is None: + return image + return foreground diff --git a/modules/ben2/ben2_model.py b/modules/ben2/ben2_model.py new file mode 100644 index 000000000..fe38c571e --- /dev/null +++ b/modules/ben2/ben2_model.py @@ -0,0 +1,1290 @@ +import os +import math +import subprocess +import tempfile +import cv2 +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.utils.checkpoint as checkpoint +import numpy as np +from timm.models.layers import DropPath, to_2tuple, trunc_normal_ +from PIL import Image +from torchvision import transforms +from einops import rearrange + + +class Mlp(nn.Module): + """ Multilayer perceptron.""" + + def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): + super().__init__() + out_features = out_features or in_features + hidden_features = hidden_features or in_features + self.fc1 = nn.Linear(in_features, hidden_features) + self.act = act_layer() + self.fc2 = nn.Linear(hidden_features, out_features) + self.drop = nn.Dropout(drop) + + def forward(self, x): + x = self.fc1(x) + x = self.act(x) + x = self.drop(x) + x = self.fc2(x) + x = self.drop(x) + return x + + +def window_partition(x, window_size): + """ + Args: + x: (B, H, W, C) + window_size (int): window size + Returns: + windows: (num_windows*B, window_size, window_size, C) + """ + B, H, W, C = x.shape + x = x.view(B, H // window_size, window_size, W // window_size, window_size, C) + windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C) + return windows + + +def window_reverse(windows, window_size, H, W): + """ + Args: + windows: (num_windows*B, window_size, window_size, C) + window_size (int): Window size + H (int): Height of image + W (int): Width of image + Returns: + x: (B, H, W, C) + """ + B = int(windows.shape[0] / (H * W / window_size / window_size)) + x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1) + x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1) + return x + + +class WindowAttention(nn.Module): + """ Window based multi-head self attention (W-MSA) module with relative position bias. + It supports both of shifted and non-shifted window. + Args: + dim (int): Number of input channels. + window_size (tuple[int]): The height and width of the window. + num_heads (int): Number of attention heads. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set + attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0 + proj_drop (float, optional): Dropout ratio of output. Default: 0.0 + """ + + def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.): + + super().__init__() + self.dim = dim + self.window_size = window_size # Wh, Ww + self.num_heads = num_heads + head_dim = dim // num_heads + self.scale = qk_scale or head_dim ** -0.5 + + # define a parameter table of relative position bias + self.relative_position_bias_table = nn.Parameter( + torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)) # 2*Wh-1 * 2*Ww-1, nH + + # get pair-wise relative position index for each token inside the window + coords_h = torch.arange(self.window_size[0]) + coords_w = torch.arange(self.window_size[1]) + coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww + coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww + relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww + relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2 + relative_coords[:, :, 0] += self.window_size[0] - 1 # shift to start from 0 + relative_coords[:, :, 1] += self.window_size[1] - 1 + relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1 + relative_position_index = relative_coords.sum(-1) # Wh*Ww, Wh*Ww + self.register_buffer("relative_position_index", relative_position_index) + + self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) + self.attn_drop = nn.Dropout(attn_drop) + self.proj = nn.Linear(dim, dim) + self.proj_drop = nn.Dropout(proj_drop) + + trunc_normal_(self.relative_position_bias_table, std=.02) + self.softmax = nn.Softmax(dim=-1) + + def forward(self, x, mask=None): + """ Forward function. + Args: + x: input features with shape of (num_windows*B, N, C) + mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None + """ + B_, N, C = x.shape + qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) + q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple) + + q = q * self.scale + attn = q @ k.transpose(-2, -1) + + relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view( + self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1) # Wh*Ww,Wh*Ww,nH + relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww + attn = attn + relative_position_bias.unsqueeze(0) + + if mask is not None: + nW = mask.shape[0] + attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0) + attn = attn.view(-1, self.num_heads, N, N) + attn = self.softmax(attn) + else: + attn = self.softmax(attn) + + attn = self.attn_drop(attn) + + x = (attn @ v).transpose(1, 2).reshape(B_, N, C) + x = self.proj(x) + x = self.proj_drop(x) + return x + + +class SwinTransformerBlock(nn.Module): + """ Swin Transformer Block. + Args: + dim (int): Number of input channels. + num_heads (int): Number of attention heads. + window_size (int): Window size. + shift_size (int): Shift size for SW-MSA. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float, optional): Stochastic depth rate. Default: 0.0 + act_layer (nn.Module, optional): Activation layer. Default: nn.GELU + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + + def __init__(self, dim, num_heads, window_size=7, shift_size=0, + mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0., + act_layer=nn.GELU, norm_layer=nn.LayerNorm): + super().__init__() + self.dim = dim + self.num_heads = num_heads + self.window_size = window_size + self.shift_size = shift_size + self.mlp_ratio = mlp_ratio + assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size" + + self.norm1 = norm_layer(dim) + self.attn = WindowAttention( + dim, window_size=to_2tuple(self.window_size), num_heads=num_heads, + qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop) + + self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() + self.norm2 = norm_layer(dim) + mlp_hidden_dim = int(dim * mlp_ratio) + self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) + + self.H = None + self.W = None + + def forward(self, x, mask_matrix): + """ Forward function. + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + mask_matrix: Attention mask for cyclic shift. + """ + B, L, C = x.shape + H, W = self.H, self.W + assert L == H * W, "input feature has wrong size" + + shortcut = x + x = self.norm1(x) + x = x.view(B, H, W, C) + + # pad feature maps to multiples of window size + pad_l = pad_t = 0 + pad_r = (self.window_size - W % self.window_size) % self.window_size + pad_b = (self.window_size - H % self.window_size) % self.window_size + x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b)) + _, Hp, Wp, _ = x.shape + + # cyclic shift + if self.shift_size > 0: + shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) + attn_mask = mask_matrix + else: + shifted_x = x + attn_mask = None + + # partition windows + x_windows = window_partition(shifted_x, self.window_size) # nW*B, window_size, window_size, C + x_windows = x_windows.view(-1, self.window_size * self.window_size, C) # nW*B, window_size*window_size, C + + # W-MSA/SW-MSA + attn_windows = self.attn(x_windows, mask=attn_mask) # nW*B, window_size*window_size, C + + # merge windows + attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) + shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp) # B H' W' C + + # reverse cyclic shift + if self.shift_size > 0: + x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)) + else: + x = shifted_x + + if pad_r > 0 or pad_b > 0: + x = x[:, :H, :W, :].contiguous() + + x = x.view(B, H * W, C) + + # FFN + x = shortcut + self.drop_path(x) + x = x + self.drop_path(self.mlp(self.norm2(x))) + + return x + + +class PatchMerging(nn.Module): + """ Patch Merging Layer + Args: + dim (int): Number of input channels. + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + """ + def __init__(self, dim, norm_layer=nn.LayerNorm): + super().__init__() + self.dim = dim + self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False) + self.norm = norm_layer(4 * dim) + + def forward(self, x, H, W): + """ Forward function. + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + """ + B, L, C = x.shape + assert L == H * W, "input feature has wrong size" + + x = x.view(B, H, W, C) + + # padding + pad_input = (H % 2 == 1) or (W % 2 == 1) + if pad_input: + x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2)) + + x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C + x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C + x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C + x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C + x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C + x = x.view(B, -1, 4 * C) # B H/2*W/2 4*C + + x = self.norm(x) + x = self.reduction(x) + + return x + + +class BasicLayer(nn.Module): + """ A basic Swin Transformer layer for one stage. + Args: + dim (int): Number of feature channels + depth (int): Depths of this stage. + num_heads (int): Number of attention head. + window_size (int): Local window size. Default: 7. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. + qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set. + drop (float, optional): Dropout rate. Default: 0.0 + attn_drop (float, optional): Attention dropout rate. Default: 0.0 + drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0 + norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm + downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + """ + + def __init__(self, + dim, + depth, + num_heads, + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + drop=0., + attn_drop=0., + drop_path=0., + norm_layer=nn.LayerNorm, + downsample=None, + use_checkpoint=False): + super().__init__() + self.window_size = window_size + self.shift_size = window_size // 2 + self.depth = depth + self.use_checkpoint = use_checkpoint + + # build blocks + self.blocks = nn.ModuleList([ + SwinTransformerBlock( + dim=dim, + num_heads=num_heads, + window_size=window_size, + shift_size=0 if (i % 2 == 0) else window_size // 2, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop, + attn_drop=attn_drop, + drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path, + norm_layer=norm_layer) + for i in range(depth)]) + + # patch merging layer + if downsample is not None: + self.downsample = downsample(dim=dim, norm_layer=norm_layer) + else: + self.downsample = None + + def forward(self, x, H, W): + """ Forward function. + Args: + x: Input feature, tensor size (B, H*W, C). + H, W: Spatial resolution of the input feature. + """ + + # calculate attention mask for SW-MSA + Hp = int(np.ceil(H / self.window_size)) * self.window_size + Wp = int(np.ceil(W / self.window_size)) * self.window_size + img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device) # 1 Hp Wp 1 + h_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + w_slices = (slice(0, -self.window_size), + slice(-self.window_size, -self.shift_size), + slice(-self.shift_size, None)) + cnt = 0 + for h in h_slices: + for w in w_slices: + img_mask[:, h, w, :] = cnt + cnt += 1 + + mask_windows = window_partition(img_mask, self.window_size) # nW, window_size, window_size, 1 + mask_windows = mask_windows.view(-1, self.window_size * self.window_size) + attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2) + attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0)) + + for blk in self.blocks: + blk.H, blk.W = H, W + if self.use_checkpoint: + x = checkpoint.checkpoint(blk, x, attn_mask) + else: + x = blk(x, attn_mask) + if self.downsample is not None: + x_down = self.downsample(x, H, W) + Wh, Ww = (H + 1) // 2, (W + 1) // 2 + return x, H, W, x_down, Wh, Ww + else: + return x, H, W, x, H, W + + +class PatchEmbed(nn.Module): + """ Image to Patch Embedding + Args: + patch_size (int): Patch token size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + norm_layer (nn.Module, optional): Normalization layer. Default: None + """ + + def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None): + super().__init__() + patch_size = to_2tuple(patch_size) + self.patch_size = patch_size + + self.in_chans = in_chans + self.embed_dim = embed_dim + + self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) + if norm_layer is not None: + self.norm = norm_layer(embed_dim) + else: + self.norm = None + + def forward(self, x): + """Forward function.""" + # padding + _, _, H, W = x.size() + if W % self.patch_size[1] != 0: + x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1])) + if H % self.patch_size[0] != 0: + x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0])) + + x = self.proj(x) # B C Wh Ww + if self.norm is not None: + Wh, Ww = x.size(2), x.size(3) + x = x.flatten(2).transpose(1, 2) + x = self.norm(x) + x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww) + + return x + + +class SwinTransformer(nn.Module): + """ Swin Transformer backbone. + A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` - + https://arxiv.org/pdf/2103.14030 + Args: + pretrain_img_size (int): Input image size for training the pretrained model, + used in absolute postion embedding. Default 224. + patch_size (int | tuple(int)): Patch size. Default: 4. + in_chans (int): Number of input image channels. Default: 3. + embed_dim (int): Number of linear projection output channels. Default: 96. + depths (tuple[int]): Depths of each Swin Transformer stage. + num_heads (tuple[int]): Number of attention head of each stage. + window_size (int): Window size. Default: 7. + mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4. + qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True + qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. + drop_rate (float): Dropout rate. + attn_drop_rate (float): Attention dropout rate. Default: 0. + drop_path_rate (float): Stochastic depth rate. Default: 0.2. + norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm. + ape (bool): If True, add absolute position embedding to the patch embedding. Default: False. + patch_norm (bool): If True, add normalization after patch embedding. Default: True. + out_indices (Sequence[int]): Output from which stages. + frozen_stages (int): Stages to be frozen (stop grad and set eval mode). + -1 means not freezing any parameters. + use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False. + """ + + def __init__(self, + pretrain_img_size=224, + patch_size=4, + in_chans=3, + embed_dim=96, + depths=[2, 2, 6, 2], + num_heads=[3, 6, 12, 24], + window_size=7, + mlp_ratio=4., + qkv_bias=True, + qk_scale=None, + drop_rate=0., + attn_drop_rate=0., + drop_path_rate=0.2, + norm_layer=nn.LayerNorm, + ape=False, + patch_norm=True, + out_indices=(0, 1, 2, 3), + frozen_stages=-1, + use_checkpoint=False): + super().__init__() + + self.pretrain_img_size = pretrain_img_size + self.num_layers = len(depths) + self.embed_dim = embed_dim + self.ape = ape + self.patch_norm = patch_norm + self.out_indices = out_indices + self.frozen_stages = frozen_stages + + # split image into non-overlapping patches + self.patch_embed = PatchEmbed( + patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim, + norm_layer=norm_layer if self.patch_norm else None) + + # absolute position embedding + if self.ape: + pretrain_img_size = to_2tuple(pretrain_img_size) + patch_size = to_2tuple(patch_size) + patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]] + + self.absolute_pos_embed = nn.Parameter(torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1])) + trunc_normal_(self.absolute_pos_embed, std=.02) + + self.pos_drop = nn.Dropout(p=drop_rate) + + # stochastic depth + dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] # stochastic depth decay rule + + # build layers + self.layers = nn.ModuleList() + for i_layer in range(self.num_layers): + layer = BasicLayer( + dim=int(embed_dim * 2 ** i_layer), + depth=depths[i_layer], + num_heads=num_heads[i_layer], + window_size=window_size, + mlp_ratio=mlp_ratio, + qkv_bias=qkv_bias, + qk_scale=qk_scale, + drop=drop_rate, + attn_drop=attn_drop_rate, + drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])], + norm_layer=norm_layer, + downsample=PatchMerging if (i_layer < self.num_layers - 1) else None, + use_checkpoint=use_checkpoint) + self.layers.append(layer) + + num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)] + self.num_features = num_features + + # add a norm layer for each output + for i_layer in out_indices: + layer = norm_layer(num_features[i_layer]) + layer_name = f'norm{i_layer}' + self.add_module(layer_name, layer) + + self._freeze_stages() + + def _freeze_stages(self): + if self.frozen_stages >= 0: + self.patch_embed.eval() + for param in self.patch_embed.parameters(): + param.requires_grad = False + + if self.frozen_stages >= 1 and self.ape: + self.absolute_pos_embed.requires_grad = False + + if self.frozen_stages >= 2: + self.pos_drop.eval() + for i in range(0, self.frozen_stages - 1): + m = self.layers[i] + m.eval() + for param in m.parameters(): + param.requires_grad = False + + + def forward(self, x): + + x = self.patch_embed(x) + + Wh, Ww = x.size(2), x.size(3) + if self.ape: + # interpolate the position embedding to the corresponding size + absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic') + x = x + absolute_pos_embed # B Wh*Ww C + + outs = [x.contiguous()] + x = x.flatten(2).transpose(1, 2) + x = self.pos_drop(x) + + + for i in range(self.num_layers): + layer = self.layers[i] + x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww) + + + if i in self.out_indices: + norm_layer = getattr(self, f'norm{i}') + x_out = norm_layer(x_out) + + out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous() + outs.append(out) + + + + return tuple(outs) + + + + + + + + +def get_activation_fn(activation): + """Return an activation function given a string""" + if activation == "gelu": + return F.gelu + + raise RuntimeError(F"activation should be gelu, not {activation}.") + + +def make_cbr(in_dim, out_dim): + return nn.Sequential(nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1), nn.InstanceNorm2d(out_dim), nn.GELU()) + + +def make_cbg(in_dim, out_dim): + return nn.Sequential(nn.Conv2d(in_dim, out_dim, kernel_size=3, padding=1), nn.InstanceNorm2d(out_dim), nn.GELU()) + + +def rescale_to(x, scale_factor: float = 2, interpolation='nearest'): + return F.interpolate(x, scale_factor=scale_factor, mode=interpolation) + + +def resize_as(x, y, interpolation='bilinear'): + return F.interpolate(x, size=y.shape[-2:], mode=interpolation) + + +def image2patches(x): + """b c (hg h) (wg w) -> (hg wg b) c h w""" + x = rearrange(x, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2 ) + return x + + +def patches2image(x): + """(hg wg b) c h w -> b c (hg h) (wg w)""" + x = rearrange(x, '(hg wg b) c h w -> b c (hg h) (wg w)', hg=2, wg=2) + return x + + + +class PositionEmbeddingSine: + def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None): + super().__init__() + self.num_pos_feats = num_pos_feats + self.temperature = temperature + self.normalize = normalize + if scale is not None and normalize is False: + raise ValueError("normalize should be True if scale is passed") + if scale is None: + scale = 2 * math.pi + self.scale = scale + self.dim_t = torch.arange(0, self.num_pos_feats, dtype=torch.float32) + + def __call__(self, b, h, w): + device = self.dim_t.device + mask = torch.zeros([b, h, w], dtype=torch.bool, device=device) + assert mask is not None + not_mask = ~mask + y_embed = not_mask.cumsum(dim=1, dtype=torch.float32) + x_embed = not_mask.cumsum(dim=2, dtype=torch.float32) + if self.normalize: + eps = 1e-6 + y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale + x_embed = (x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * self.scale + + dim_t = self.temperature ** (2 * (self.dim_t.to(device) // 2) / self.num_pos_feats) + pos_x = x_embed[:, :, :, None] / dim_t + pos_y = y_embed[:, :, :, None] / dim_t + + pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3) + pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3) + + return torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) + + + +class MCLM(nn.Module): + def __init__(self, d_model, num_heads, pool_ratios=[1, 4, 8]): + super(MCLM, self).__init__() + self.attention = nn.ModuleList([ + nn.MultiheadAttention(d_model, num_heads, dropout=0.1), + nn.MultiheadAttention(d_model, num_heads, dropout=0.1), + nn.MultiheadAttention(d_model, num_heads, dropout=0.1), + nn.MultiheadAttention(d_model, num_heads, dropout=0.1), + nn.MultiheadAttention(d_model, num_heads, dropout=0.1) + ]) + + self.linear1 = nn.Linear(d_model, d_model * 2) + self.linear2 = nn.Linear(d_model * 2, d_model) + self.linear3 = nn.Linear(d_model, d_model * 2) + self.linear4 = nn.Linear(d_model * 2, d_model) + self.norm1 = nn.LayerNorm(d_model) + self.norm2 = nn.LayerNorm(d_model) + self.dropout = nn.Dropout(0.1) + self.dropout1 = nn.Dropout(0.1) + self.dropout2 = nn.Dropout(0.1) + self.activation = get_activation_fn('gelu') + self.pool_ratios = pool_ratios + self.p_poses = [] + self.g_pos = None + self.positional_encoding = PositionEmbeddingSine(num_pos_feats=d_model // 2, normalize=True) + + def forward(self, l, g): + """ + l: 4,c,h,w + g: 1,c,h,w + """ + self.p_poses = [] + self.g_pos = None + _b, _c, h, w = l.size() + # 4,c,h,w -> 1,c,2h,2w + concated_locs = rearrange(l, '(hg wg b) c h w -> b c (hg h) (wg w)', hg=2, wg=2) + + pools = [] + for pool_ratio in self.pool_ratios: + # b,c,h,w + tgt_hw = (round(h / pool_ratio), round(w / pool_ratio)) + pool = F.adaptive_avg_pool2d(concated_locs, tgt_hw) + pools.append(rearrange(pool, 'b c h w -> (h w) b c')) + if self.g_pos is None: + pos_emb = self.positional_encoding(pool.shape[0], pool.shape[2], pool.shape[3]) + pos_emb = rearrange(pos_emb, 'b c h w -> (h w) b c') + self.p_poses.append(pos_emb) + pools = torch.cat(pools, 0) + if self.g_pos is None: + self.p_poses = torch.cat(self.p_poses, dim=0) + pos_emb = self.positional_encoding(g.shape[0], g.shape[2], g.shape[3]) + self.g_pos = rearrange(pos_emb, 'b c h w -> (h w) b c') + + device = pools.device + self.p_poses = self.p_poses.to(device) + self.g_pos = self.g_pos.to(device) + + + # attention between glb (q) & multisensory concated-locs (k,v) + g_hw_b_c = rearrange(g, 'b c h w -> (h w) b c') + + + g_hw_b_c = g_hw_b_c + self.dropout1(self.attention[0](g_hw_b_c + self.g_pos, pools + self.p_poses, pools)[0]) + g_hw_b_c = self.norm1(g_hw_b_c) + g_hw_b_c = g_hw_b_c + self.dropout2(self.linear2(self.dropout(self.activation(self.linear1(g_hw_b_c)).clone()))) + g_hw_b_c = self.norm2(g_hw_b_c) + + # attention between origin locs (q) & freashed glb (k,v) + l_hw_b_c = rearrange(l, "b c h w -> (h w) b c") + _g_hw_b_c = rearrange(g_hw_b_c, '(h w) b c -> h w b c', h=h, w=w) + _g_hw_b_c = rearrange(_g_hw_b_c, "(ng h) (nw w) b c -> (h w) (ng nw b) c", ng=2, nw=2) + outputs_re = [] + for i, (_l, _g) in enumerate(zip(l_hw_b_c.chunk(4, dim=1), _g_hw_b_c.chunk(4, dim=1))): + outputs_re.append(self.attention[i + 1](_l, _g, _g)[0]) # (h w) 1 c + outputs_re = torch.cat(outputs_re, 1) # (h w) 4 c + + l_hw_b_c = l_hw_b_c + self.dropout1(outputs_re) + l_hw_b_c = self.norm1(l_hw_b_c) + l_hw_b_c = l_hw_b_c + self.dropout2(self.linear4(self.dropout(self.activation(self.linear3(l_hw_b_c)).clone()))) + l_hw_b_c = self.norm2(l_hw_b_c) + + l = torch.cat((l_hw_b_c, g_hw_b_c), 1) # hw,b(5),c + return rearrange(l, "(h w) b c -> b c h w", h=h, w=w) ## (5,c,h*w) + + + + + + + + + +class MCRM(nn.Module): + def __init__(self, d_model, num_heads, pool_ratios=[4, 8, 16], h=None): # pylint: disable=unused-argument + super(MCRM, self).__init__() + self.attention = nn.ModuleList([ + nn.MultiheadAttention(d_model, num_heads, dropout=0.1), + nn.MultiheadAttention(d_model, num_heads, dropout=0.1), + nn.MultiheadAttention(d_model, num_heads, dropout=0.1), + nn.MultiheadAttention(d_model, num_heads, dropout=0.1) + ]) + self.linear3 = nn.Linear(d_model, d_model * 2) + self.linear4 = nn.Linear(d_model * 2, d_model) + self.norm1 = nn.LayerNorm(d_model) + self.norm2 = nn.LayerNorm(d_model) + self.dropout = nn.Dropout(0.1) + self.dropout1 = nn.Dropout(0.1) + self.dropout2 = nn.Dropout(0.1) + self.sigmoid = nn.Sigmoid() + self.activation = get_activation_fn('gelu') + self.sal_conv = nn.Conv2d(d_model, 1, 1) + self.pool_ratios = pool_ratios + + def forward(self, x): + # device = x.device + _b, c, h, w = x.size() + loc, glb = x.split([4, 1], dim=0) # 4,c,h,w; 1,c,h,w + + patched_glb = rearrange(glb, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2) + + token_attention_map = self.sigmoid(self.sal_conv(glb)) + token_attention_map = F.interpolate(token_attention_map, size=patches2image(loc).shape[-2:], mode='nearest') + loc = loc * rearrange(token_attention_map, 'b c (hg h) (wg w) -> (hg wg b) c h w', hg=2, wg=2) + + pools = [] + for pool_ratio in self.pool_ratios: + tgt_hw = (round(h / pool_ratio), round(w / pool_ratio)) + pool = F.adaptive_avg_pool2d(patched_glb, tgt_hw) + pools.append(rearrange(pool, 'nl c h w -> nl c (h w)')) # nl(4),c,hw + + pools = rearrange(torch.cat(pools, 2), "nl c nphw -> nl nphw 1 c") + loc_ = rearrange(loc, 'nl c h w -> nl (h w) 1 c') + + outputs = [] + for i, q in enumerate(loc_.unbind(dim=0)): # traverse all local patches + v = pools[i] + k = v + outputs.append(self.attention[i](q, k, v)[0]) + + outputs = torch.cat(outputs, 1) + src = loc.view(4, c, -1).permute(2, 0, 1) + self.dropout1(outputs) + src = self.norm1(src) + src = src + self.dropout2(self.linear4(self.dropout(self.activation(self.linear3(src)).clone()))) + src = self.norm2(src) + src = src.permute(1, 2, 0).reshape(4, c, h, w) # freshed loc + glb = glb + F.interpolate(patches2image(src), size=glb.shape[-2:], mode='nearest') # freshed glb + + return torch.cat((src, glb), 0), token_attention_map + + + +class BEN_Base(nn.Module): + def __init__(self): + super().__init__() + + self.backbone = SwinTransformer(embed_dim=128, depths=[2, 2, 18, 2], num_heads=[4, 8, 16, 32], window_size=12) + emb_dim = 128 + self.sideout5 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) + self.sideout4 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) + self.sideout3 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) + self.sideout2 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) + self.sideout1 = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) + + self.output5 = make_cbr(1024, emb_dim) + self.output4 = make_cbr(512, emb_dim) + self.output3 = make_cbr(256, emb_dim) + self.output2 = make_cbr(128, emb_dim) + self.output1 = make_cbr(128, emb_dim) + + self.multifieldcrossatt = MCLM(emb_dim, 1, [1, 4, 8]) + self.conv1 = make_cbr(emb_dim, emb_dim) + self.conv2 = make_cbr(emb_dim, emb_dim) + self.conv3 = make_cbr(emb_dim, emb_dim) + self.conv4 = make_cbr(emb_dim, emb_dim) + self.dec_blk1 = MCRM(emb_dim, 1, [2, 4, 8]) + self.dec_blk2 = MCRM(emb_dim, 1, [2, 4, 8]) + self.dec_blk3 = MCRM(emb_dim, 1, [2, 4, 8]) + self.dec_blk4 = MCRM(emb_dim, 1, [2, 4, 8]) + + self.insmask_head = nn.Sequential( + nn.Conv2d(emb_dim, 384, kernel_size=3, padding=1), + nn.InstanceNorm2d(384), + nn.GELU(), + nn.Conv2d(384, 384, kernel_size=3, padding=1), + nn.InstanceNorm2d(384), + nn.GELU(), + nn.Conv2d(384, emb_dim, kernel_size=3, padding=1) + ) + + self.shallow = nn.Sequential(nn.Conv2d(3, emb_dim, kernel_size=3, padding=1)) + self.upsample1 = make_cbg(emb_dim, emb_dim) + self.upsample2 = make_cbg(emb_dim, emb_dim) + self.output = nn.Sequential(nn.Conv2d(emb_dim, 1, kernel_size=3, padding=1)) + + for m in self.modules(): + if isinstance(m, nn.GELU) or isinstance(m, nn.Dropout): + m.inplace = True + + + @torch.inference_mode() + @torch.autocast(device_type="cuda",dtype=torch.float16) + def forward(self, x): + real_batch = x.size(0) + + shallow_batch = self.shallow(x) + glb_batch = rescale_to(x, scale_factor=0.5, interpolation='bilinear') + + + + final_input = None + for i in range(real_batch): + start = i * 4 + end = (i + 1) * 4 + loc_batch = image2patches(x[i,:,:,:].unsqueeze(dim=0)) + input_ = torch.cat((loc_batch, glb_batch[i,:,:,:].unsqueeze(dim=0)), dim=0) + if final_input is None: + final_input= input_ + else: + final_input = torch.cat((final_input, input_), dim=0) + + features = self.backbone(final_input) + outputs = [] + for i in range(real_batch): + + start = i * 5 + end = (i + 1) * 5 + f4 = features[4][start:end, :, :, :] # shape: [5, C, H, W] + f3 = features[3][start:end, :, :, :] + f2 = features[2][start:end, :, :, :] + f1 = features[1][start:end, :, :, :] + f0 = features[0][start:end, :, :, :] + e5 = self.output5(f4) + e4 = self.output4(f3) + e3 = self.output3(f2) + e2 = self.output2(f1) + e1 = self.output1(f0) + loc_e5, glb_e5 = e5.split([4, 1], dim=0) + e5 = self.multifieldcrossatt(loc_e5, glb_e5) # (4,128,16,16) + + + e4, _tokenattmap4 = self.dec_blk4(e4 + resize_as(e5, e4)) + e4 = self.conv4(e4) + e3, _tokenattmap3 = self.dec_blk3(e3 + resize_as(e4, e3)) + e3 = self.conv3(e3) + e2, _tokenattmap2 = self.dec_blk2(e2 + resize_as(e3, e2)) + e2 = self.conv2(e2) + e1, _tokenattmap1 = self.dec_blk1(e1 + resize_as(e2, e1)) + e1 = self.conv1(e1) + + loc_e1, glb_e1 = e1.split([4, 1], dim=0) + + output1_cat = patches2image(loc_e1) # (1,128,256,256) + + # add glb feat in + output1_cat = output1_cat + resize_as(glb_e1, output1_cat) + # merge + final_output = self.insmask_head(output1_cat) # (1,128,256,256) + # shallow feature merge + shallow = shallow_batch[i,:,:,:].unsqueeze(dim=0) + final_output = final_output + resize_as(shallow, final_output) + final_output = self.upsample1(rescale_to(final_output)) + final_output = rescale_to(final_output + resize_as(shallow, final_output)) + final_output = self.upsample2(final_output) + final_output = self.output(final_output) + mask = final_output.sigmoid() + outputs.append(mask) + + return torch.cat(outputs, dim=0) + + + def loadcheckpoints(self,model_path): + model_dict = torch.load(model_path, map_location="cpu", weights_only=True) + self.load_state_dict(model_dict['model_state_dict'], strict=True) + del model_path + + def inference(self,image,refine_foreground=False): + # image = ImageOps.exif_transpose(image) + if isinstance(image, Image.Image): + image, h, w,original_image = rgb_loader_refiner(image) + if torch.cuda.is_available(): + + img_tensor = img_transform(image).unsqueeze(0).to(next(self.parameters()).device) + else: + img_tensor = img_transform32(image).unsqueeze(0).to(next(self.parameters()).device) + + with torch.no_grad(): + res = self.forward(img_tensor) + + # Show Results + if refine_foreground: + + pred_pil = transforms.ToPILImage()(res.squeeze()) + image_masked = refine_foreground_process(original_image, pred_pil) + image_masked.putalpha(pred_pil.resize(original_image.size)) + return image_masked + + else: + alpha = postprocess_image(res, im_size=[w,h]) + pred_pil = transforms.ToPILImage()(alpha) + mask = pred_pil.resize(original_image.size) + original_image.putalpha(mask) + # mask = Image.fromarray(alpha) + + return original_image + + + else: + foregrounds = [] + for batch in image: + image, h, w,original_image = rgb_loader_refiner(batch) + if torch.cuda.is_available(): + + img_tensor = img_transform(image).unsqueeze(0).to(next(self.parameters()).device) + else: + img_tensor = img_transform32(image).unsqueeze(0).to(next(self.parameters()).device) + + with torch.no_grad(): + res = self.forward(img_tensor) + + if refine_foreground: + + pred_pil = transforms.ToPILImage()(res.squeeze()) + image_masked = refine_foreground_process(original_image, pred_pil) + image_masked.putalpha(pred_pil.resize(original_image.size)) + + foregrounds.append(image_masked) + else: + alpha = postprocess_image(res, im_size=[w,h]) + pred_pil = transforms.ToPILImage()(alpha) + mask = pred_pil.resize(original_image.size) + original_image.putalpha(mask) + # mask = Image.fromarray(alpha) + foregrounds.append(original_image) + + return foregrounds + + + + + def segment_video(self, video_path, output_path="./", fps=0, refine_foreground=False, batch=1, print_frames_processed=True, webm = False, rgb_value= (0, 255, 0)): + """ + Segments the given video to extract the foreground (with alpha) from each frame + and saves the result as either a WebM video (with alpha channel) or MP4 (with a + color background). + + Args: + video_path (str): + Path to the input video file. + + output_path (str, optional): + Directory (or full path) where the output video and/or files will be saved. + Defaults to "./". + + fps (int, optional): + The frames per second (FPS) to use for the output video. If 0 (default), the + original FPS of the input video is used. Otherwise, overrides it. + + refine_foreground (bool, optional): + Whether to run an additional “refine foreground” process on each frame. + Defaults to False. + + batch (int, optional): + Number of frames to process at once (inference batch size). Large batch sizes + may require more GPU memory. Defaults to 1. + + print_frames_processed (bool, optional): + If True (default), prints progress (how many frames have been processed) to + the console. + + webm (bool, optional): + If True (default), exports a WebM video with alpha channel (VP9 / yuva420p). + If False, exports an MP4 video composited over a solid color background. + + rgb_value (tuple, optional): + The RGB background color (e.g., green screen) used to composite frames when + saving to MP4. Defaults to (0, 255, 0). + + Returns: + None. Writes the output video(s) to disk in the specified format. + """ + cap = cv2.VideoCapture(video_path) + if not cap.isOpened(): + raise IOError(f"Cannot open video: {video_path}") + + original_fps = cap.get(cv2.CAP_PROPFPS) + original_fps = 30 if original_fps == 0 else original_fps + fps = original_fps if fps == 0 else fps + + ret, first_frame = cap.read() + if not ret: + raise ValueError("No frames found in the video.") + _height, _width = first_frame.shape[:2] + cap.set(cv2.CAP_PROP_POSFRAMES, 0) + + foregrounds = [] + frame_idx = 0 + processed_count = 0 + batch_frames = [] + total_frames = int(cap.get(cv2.CAP_PROPFRAME_COUNT)) + + while True: + ret, frame = cap.read() + if not ret: + if batch_frames: + batch_results = self.inference(batch_frames, refine_foreground) + if isinstance(batch_results, Image.Image): + foregrounds.append(batch_results) + else: + foregrounds.extend(batch_results) + if print_frames_processed: + print(f"Processed frames {frame_idx-len(batch_frames)+1} to {frame_idx} of {total_frames}") + break + + # Process every frame instead of using intervals + frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) + pil_frame = Image.fromarray(frame_rgb) + batch_frames.append(pil_frame) + if len(batch_frames) == batch: + batch_results = self.inference(batch_frames, refine_foreground) + if isinstance(batch_results, Image.Image): + foregrounds.append(batch_results) + else: + foregrounds.extend(batch_results) + if print_frames_processed: + print(f"Processed frames {frame_idx-batch+1} to {frame_idx} of {total_frames}") + batch_frames = [] + processed_count += batch + + frame_idx += 1 + + + if webm: + alpha_webm_path = os.path.join(output_path, "foreground.webm") + pil_images_to_webm_alpha(foregrounds, alpha_webm_path, fps=original_fps) + + else: + cap.release() + fg_output = os.path.join(output_path, 'foreground.mp4') + pil_images_to_mp4(foregrounds, fg_output, fps=original_fps,rgb_value=rgb_value) + cv2.destroyAllWindows() + try: + fg_audio_output = os.path.join(output_path, 'foreground_output_with_audio.mp4') + add_audio_to_video(fg_output, video_path, fg_audio_output) + except Exception as e: + print("No audio found in the original video") + print(e) + + +def rgb_loader_refiner( original_image): + h, w = original_image.size + image = original_image + # Convert to RGB if necessary + if image.mode != 'RGB': + image = image.convert('RGB') + # Resize the image + image = image.resize((1024, 1024), resample=Image.Resampling.LANCZOS) + return image.convert('RGB'), h, w,original_image + + +# Define the image transformation +img_transform = transforms.Compose([ + transforms.ToTensor(), + transforms.ConvertImageDtype(torch.float16), + transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) +]) + +img_transform32 = transforms.Compose([ + transforms.ToTensor(), + transforms.ConvertImageDtype(torch.float32), + transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) +]) + + +def pil_images_to_mp4(images, output_path, fps=24, rgb_value=(0, 255, 0)): + """ + Converts an array of PIL images to an MP4 video. + Args: + images: List of PIL images + output_path: Path to save the MP4 file + fps: Frames per second (default: 24) + rgb_value: Background RGB color tuple (default: green (0, 255, 0)) + """ + if not images: + raise ValueError("No images provided to convert to MP4.") + + width, height = images[0].size + fourcc = cv2.VideoWriter_fourcc(*'mp4v') + video_writer = cv2.VideoWriter(output_path, fourcc, fps, (width, height)) + + for image in images: + # If image has alpha channel, composite onto the specified background color + if image.mode == 'RGBA': + # Create background image with specified RGB color + background = Image.new('RGB', image.size, rgb_value) + background = background.convert('RGBA') + # Composite the image onto the background + image = Image.alpha_composite(background, image) + image = image.convert('RGB') + else: + # Ensure RGB format for non-alpha images + image = image.convert('RGB') + + # Convert to OpenCV format and write + open_cv_image = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR) + video_writer.write(open_cv_image) + video_writer.release() + +def pil_images_to_webm_alpha(images, output_path, fps=30): + """ + Converts a list of PIL RGBA images to a VP9 .webm video with alpha channel. + + NOTE: Not all players will display alpha in WebM. + Browsers like Chrome/Firefox typically do support VP9 alpha. + """ + if not images: + raise ValueError("No images provided for WebM with alpha.") + + # Ensure output directory exists + os.makedirs(os.path.dirname(output_path), exist_ok=True) + + with tempfile.TemporaryDirectory() as tmpdir: + # Save frames as PNG (with alpha) + for idx, img in enumerate(images): + if img.mode != "RGBA": + img = img.convert("RGBA") + out_path = os.path.join(tmpdir, f"{idx:06d}.png") + img.save(out_path, "PNG") + + # Construct ffmpeg command + # -c:v libvpx-vp9 => VP9 encoder + # -pix_fmt yuva420p => alpha-enabled pixel format + # -auto-alt-ref 0 => helps preserve alpha frames (libvpx quirk) + ffmpeg_cmd = [ + "ffmpeg", "-y", + "-framerate", str(fps), + "-i", os.path.join(tmpdir, "%06d.png"), + "-c:v", "libvpx-vp9", + "-pix_fmt", "yuva420p", + "-auto-alt-ref", "0", + output_path + ] + + subprocess.run(ffmpeg_cmd, check=True) + + print(f"WebM with alpha saved to {output_path}") + +def add_audio_to_video(video_without_audio_path, original_video_path, output_path): + """ + Check if the original video has an audio stream. If yes, add it. If not, skip. + """ + # 1) Probe original video for audio streams + probe_command = [ + 'ffprobe', '-v', 'error', + '-select_streams', 'a:0', + '-show_entries', 'stream=index', + '-of', 'csv=p=0', + original_video_path + ] + result = subprocess.run(probe_command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False) + + # result.stdout is empty if no audio stream found + if not result.stdout.strip(): + print("No audio track found in original video, skipping audio addition.") + return + print("Audio track detected; proceeding to mux audio.") + # 2) If audio found, run ffmpeg to add it + command = [ + 'ffmpeg', '-y', + '-i', video_without_audio_path, + '-i', original_video_path, + '-c', 'copy', + '-map', '0:v:0', + '-map', '1:a:0', # we know there's an audio track now + output_path + ] + subprocess.run(command, check=True) + print(f"Audio added successfully => {output_path}") + + +### Thanks to the source: https://huggingface.co/ZhengPeng7/BiRefNet/blob/main/handler.py +def refine_foreground_process(image, mask, r=90): + if mask.size != image.size: + mask = mask.resize(image.size) + image = np.array(image) / 255.0 + mask = np.array(mask) / 255.0 + estimated_foreground = FB_blur_fusion_foreground_estimator_2(image, mask, r=r) + image_masked = Image.fromarray((estimated_foreground * 255.0).astype(np.uint8)) + return image_masked + + +def FB_blur_fusion_foreground_estimator_2(image, alpha, r=90): + # Thanks to the source: https://github.com/Photoroom/fast-foreground-estimation + alpha = alpha[:, :, None] + F, blur_B = FB_blur_fusion_foreground_estimator(image, image, image, alpha, r) + return FB_blur_fusion_foreground_estimator(image, F, blur_B, alpha, r=6)[0] + + +def FB_blur_fusion_foreground_estimator(image, F, B, alpha, r=90): + if isinstance(image, Image.Image): + image = np.array(image) / 255.0 + blurred_alpha = cv2.blur(alpha, (r, r))[:, :, None] + + blurredFA = cv2.blur(F * alpha, (r, r)) + blurredF = blurredFA / (blurred_alpha + 1e-5) + + blurred_B1A = cv2.blur(B * (1 - alpha), (r, r)) + blurred_B = blurred_B1A / ((1 - blurred_alpha) + 1e-5) + F = blurredF + alpha * \ + (image - alpha * blurredF - (1 - alpha) * blurred_B) + F = np.clip(F, 0, 1) + return F, blurred_B + + +def postprocess_image(result: torch.Tensor, im_size: list) -> np.ndarray: + result = torch.squeeze(F.interpolate(result, size=im_size, mode='bilinear'), 0) + ma = torch.max(result) + mi = torch.min(result) + result = (result - mi) / (ma - mi) + im_array = (result * 255).permute(1, 2, 0).cpu().data.numpy().astype(np.uint8) + im_array = np.squeeze(im_array) + return im_array diff --git a/modules/masking.py b/modules/masking.py index 4c436a288..bfebe8d83 100644 --- a/modules/masking.py +++ b/modules/masking.py @@ -133,12 +133,12 @@ MODELS = { 'Facebook SAM ViT Huge': 'facebook/sam-vit-huge', 'SlimSAM Uniform': 'Zigeng/SlimSAM-uniform-50', 'SlimSAM Uniform Tiny': 'Zigeng/SlimSAM-uniform-77', + 'Rembg BEN2': 'ben2', 'Rembg Silueta': 'silueta', 'Rembg U2Net': 'u2net', - 'Rembg ISNet': 'isnet', - # "u2net_human_seg", - # "isnet-general-use", - # "isnet-anime", + 'Rembg U2Net human': 'u2net_human_seg', + 'Rembg ISNet general': 'isnet-general-use', + 'Rembg ISNet anime': 'isnet-anime', } COLORMAP = ['autumn', 'bone', 'jet', 'winter', 'rainbow', 'ocean', 'summer', 'spring', 'cool', 'hsv', 'pink', 'hot', 'parula', 'magma', 'inferno', 'plasma', 'viridis', 'cividis', 'twilight', 'shifted', 'turbo', 'deepgreen'] TYPES = ['None', 'Opaque', 'Binary', 'Masked', 'Grayscale', 'Color', 'Composite'] @@ -152,12 +152,13 @@ controls = [] opts = SimpleNamespace(**{ 'model': None, 'auto_mask': 'None', + 'auto_segment': 'None', 'mask_only': False, - 'mask_blur': 0.01, - 'mask_erode': 0.01, - 'mask_dilate': 0.01, + 'mask_blur': 0, + 'mask_erode': 0, + 'mask_dilate': 0, 'seg_iou_thresh': 0.5, - 'seg_score_thresh': 0.5, + 'seg_score_thresh': 0.8, 'seg_nms_thresh': 0.5, 'seg_overlap_ratio': 0.3, 'seg_points_per_batch': 64, @@ -190,7 +191,7 @@ def init_model(selected_model: str): if opts.model != selected_model or generator is None: # sam pipeline busy = True t0 = time.time() - shared.log.debug(f'Mask segment loading: model={selected_model} path={model_path}') + shared.log.debug(f'Mask segment loading: model="{selected_model}" path={model_path}') model = SamModel.from_pretrained(model_path, cache_dir=cache_dir).to(device=devices.device) processor = SamImageProcessor.from_pretrained(model_path, cache_dir=cache_dir) generator = MaskGenerationPipeline( @@ -201,7 +202,7 @@ def init_model(selected_model: str): # output_rle_masks=False, ) devices.torch_gc() - shared.log.debug(f'Mask segment loaded: model={selected_model} path={model_path} time={time.time()-t0:.2f}s') + shared.log.debug(f'Mask segment loaded: model="{selected_model}" path={model_path} time={time.time()-t0:.2f}s') opts.model = selected_model busy = False return selected_model @@ -227,10 +228,14 @@ def run_segment(input_image: gr.Image, input_mask: np.ndarray): return outputs devices.torch_gc() i = 1 + if input_mask is None: + input_mask = np.zeros((input_image.height, input_image.width), dtype='uint8') + elif isinstance(input_mask, Image.Image): + input_mask = np.array(input_mask) combined_mask = np.zeros(input_mask.shape, dtype='uint8') input_mask_size = np.count_nonzero(input_mask) debug(f'Segment SAM: {vars(opts)}') - for mask in outputs['masks']: + for mask, score in zip(outputs['masks'], outputs['scores']): mask = mask.astype('uint8') mask_size = np.count_nonzero(mask) if mask_size == 0: @@ -245,7 +250,7 @@ def run_segment(input_image: gr.Image, input_mask: np.ndarray): continue mask = (opts.seg_topK + 1 - i) * mask * (255 // opts.seg_topK) # set grayscale intensity so we can recolor combined_mask = combined_mask + mask - debug(f'Segment mask: i={i} size={input_image.width}x{input_image.height} masked={mask_size}px overlap={overlap} score={outputs["scores"][i-1]:.2f}') + debug(f'Segment mask: i={i} size={input_image.width}x{input_image.height} masked={mask_size}px overlap={overlap} score={score:.2f}') i += 1 if i > opts.seg_topK: break @@ -260,21 +265,33 @@ def run_rembg(input_image: Image, input_mask: np.ndarray): return input_mask if "U2NET_HOME" not in os.environ: os.environ["U2NET_HOME"] = os.path.join(paths.models_path, "Rembg") - args = { - 'data': input_image, - 'only_mask': True, - 'post_process_mask': False, - 'bgcolor': None, - 'alpha_matting': False, - 'alpha_matting_foreground_threshold': 240, - 'alpha_matting_background_threshold': 10, - 'alpha_matting_erode_size': int(opts.mask_erode * 40), - 'session': rembg.new_session(opts.model), - } - mask = rembg.remove(**args) + if opts.model == 'ben2': + from modules import ben2 + args = { + 'image': input_image, + 'refine': True, + } + mask = ben2.remove(**args) + _r, _g, _b, alpha = mask.split() + mask = alpha + else: + args = { + 'data': input_image, + 'only_mask': True, + 'post_process_mask': False, + 'bgcolor': None, + 'alpha_matting': False, + 'alpha_matting_foreground_threshold': 240, + 'alpha_matting_background_threshold': 10, + 'alpha_matting_erode_size': int(opts.mask_erode * 40), + 'session': rembg.new_session(opts.model), + } + mask = rembg.remove(**args) mask = np.array(mask) - if len(input_mask.shape) > 2: - mask = cv2.cvtColor(input_mask, cv2.COLOR_RGB2GRAY) + if input_mask is None: + input_mask = np.zeros(mask.shape, dtype='uint8') + elif isinstance(input_mask, Image.Image): + input_mask = np.array(input_mask) binary_input = cv2.threshold(input_mask, 127, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU)[1] binary_output = cv2.threshold(mask, 127, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU)[1] if binary_input.shape != binary_output.shape: @@ -289,6 +306,7 @@ def run_rembg(input_image: Image, input_mask: np.ndarray): def get_mask(input_image: gr.Image, input_mask: gr.Image): + debug('Run auto-mask') # pylint: disable=protected-access t0 = time.time() if input_mask is not None: output_mask = np.array(input_mask) @@ -349,30 +367,6 @@ def outpaint(input_image: Image.Image, outpaint_type: str = 'Edge'): mask = cv2.resize(mask, (w0, h0)) mask = cv2.cvtColor(np.array(mask), cv2.COLOR_BGR2GRAY) mask = cv2.threshold(mask, 0, 255, cv2.THRESH_BINARY)[1] - """ - size = min(input_image.width, input_image.height) - if opts.mask_erode > 0: - try: - kernel = np.ones((int(opts.mask_erode * size / 4) + 1, int(opts.mask_erode * size / 4) + 1), np.uint8) - mask = cv2.erode(mask, kernel, iterations=opts.kernel_iterations) # remove noise - debug(f'Mask erode={opts.mask_erode:.3f} kernel={kernel.shape} mask={mask.shape}') - except Exception as e: - shared.log.error(f'Mask erode: {e}') - if opts.mask_dilate > 0: - try: - kernel = np.ones((int(opts.mask_dilate * size / 4) + 1, int(opts.mask_dilate * size / 4) + 1), np.uint8) - mask = cv2.dilate(mask, kernel, iterations=opts.kernel_iterations) # expand area - debug(f'Mask dilate={opts.mask_dilate:.3f} kernel={kernel.shape} mask={mask.shape}') - except Exception as e: - shared.log.error(f'Mask dilate: {e}') - if opts.mask_blur > 0: - try: - sigmax, sigmay = 1 + int(opts.mask_blur * size / 4), 1 + int(opts.mask_blur * size / 4) - mask = cv2.GaussianBlur(mask, (0, 0), sigmaX=sigmax, sigmaY=sigmay) # blur mask - debug(f'Mask blur={opts.mask_blur:.3f} x={sigmax} y={sigmay} mask={mask.shape}') - except Exception as e: - shared.log.error(f'Mask blur: {e}') - """ if outpaint_type == 'Edge': bordered = cv2.copyMakeBorder(cropped, y1, h0-y2, x1, w0-x2, cv2.BORDER_REPLICATE) bordered = cv2.resize(bordered, (w0, h0)) @@ -384,44 +378,55 @@ def outpaint(input_image: Image.Image, outpaint_type: str = 'Edge'): return image, mask -def run_mask(input_image: Image.Image, input_mask: Image.Image = None, return_type: str = None, mask_blur: int = None, mask_padding: int = None, segment_enable=True, invert=None): - fn = f'{sys._getframe(2).f_code.co_name}:{sys._getframe(1).f_code.co_name}' # pylint: disable=protected-access - debug(f'Run mask: fn={fn}') # pylint: disable=protected-access - - if input_image is None: - return input_mask - if isinstance(input_image, list): +def run_mask(input_image: Image.Image, input_mask: Image.Image = None, return_type: str = None, mask_blur: int = None, mask_padding: int = None, invert=None): + if isinstance(input_image, list) and len(input_image) > 0: input_image = input_image[0] - if isinstance(input_image, dict): + elif isinstance(input_image, dict): input_mask = input_image.get('mask', None) input_image = input_image.get('image', None) - if input_image is None: + elif isinstance(input_image, np.ndarray): + input_image = Image.fromarray(input_image) + elif isinstance(input_image, Image.Image): + pass + else: return input_mask - # t0 = time.time() - input_mask = get_mask(input_image, input_mask) # perform optional auto-masking - if input_mask is None: - return None + fn = f'{sys._getframe(2).f_code.co_name}:{sys._getframe(1).f_code.co_name}' # pylint: disable=protected-access + debug(f'Run mask: fn={fn}') # pylint: disable=protected-access + debug(f'Run mask: opts={opts}') # pylint: disable=protected-access size = min(input_image.width, input_image.height) - if mask_blur is not None or mask_padding is not None: - debug(f'Mask args legacy: blur={mask_blur} padding={mask_padding}') if invert is not None: opts.invert = invert - if mask_blur is not None: # compatibility with old img2img values which uses px values - opts.mask_blur = round(4 * mask_blur / size, 3) - if mask_padding is not None: # compatibility with old img2img values which uses px values - opts.mask_dilate = 4 * mask_padding / size - if opts.model is None or not segment_enable: - mask = input_mask - elif generator is None: - mask = run_rembg(input_image, input_mask) - else: - mask = run_segment(input_image, input_mask) + # set legacy mask args + if mask_blur is not None or mask_padding is not None: + debug(f'Mask args legacy: blur={mask_blur} padding={mask_padding}') + if mask_blur is not None: # compatibility with old img2img values which uses px values + opts.mask_blur = round(4 * mask_blur / size, 3) + if mask_padding is not None: # compatibility with old img2img values which uses px values + size = min(input_image.width, input_image.height) + opts.mask_dilate = 4 * mask_padding / size + + + # optional auto-masking and auto-segmentation + mask = input_mask + if opts.auto_mask is not None and opts.auto_mask != 'None': + mask = get_mask(input_image, input_mask) # perform optional auto-masking + elif opts.auto_segment is not None and opts.auto_segment != 'None': + init_model(opts.auto_segment) + if generator is not None: + mask = run_segment(input_image, input_mask) + else: + mask = run_rembg(input_image, input_mask) + elif isinstance(mask, Image.Image): + mask = np.array(mask) + + # early exit if no input mask or auto-mask + if mask is None: + return None mask = cv2.resize(mask, (input_image.width, input_image.height), interpolation=cv2.INTER_LANCZOS4) - # shared.log.trace(f'Mask shape={mask.shape} opts={opts} fn={fn}') if opts.mask_erode > 0: try: kernel = np.ones((int(opts.mask_erode * size / 4) + 1, int(opts.mask_erode * size / 4) + 1), np.uint8) @@ -448,11 +453,6 @@ def run_mask(input_image: Image.Image, input_mask: Image.Image = None, return_ty return_type = return_type or opts.preview_type - # mask_size = np.count_nonzero(mask) - # total_size = np.prod(mask.shape) - # area_size = np.count_nonzero(mask) - # t1 = time.time() - # shared.log.debug(f'Mask: size={input_image.width}x{input_image.height} masked={mask_size}px area={area_size/total_size:.2f} auto={opts.auto_mask} blur={opts.mask_blur:.3f} erode={opts.mask_erode:.3f} dilate={opts.mask_dilate:.3f} type={return_type} time={t1-t0:.2f}') if return_type == 'None': return input_mask elif return_type == 'Opaque': @@ -518,42 +518,44 @@ def create_segment_ui(): opts.seg_live = args[0] opts.mask_only = args[1] opts.invert = args[2] - opts.mask_blur = args[3] + opts.mask_dilate = args[3] opts.mask_erode = args[4] - opts.mask_dilate = args[5] - opts.auto_mask = args[6] - opts.seg_score_thresh = args[7] - opts.seg_iou_thresh = args[8] - opts.seg_nms_thresh = args[9] - opts.preview_type = args[10] - opts.seg_colormap = args[11] + opts.mask_blur = args[5] + opts.seg_score_thresh = args[6] + opts.auto_mask = args[7] + opts.auto_segment = args[8] + opts.seg_iou_thresh = args[9] + opts.seg_nms_thresh = args[10] + opts.preview_type = args[11] + opts.seg_colormap = args[12] global btn_mask, btn_lama # pylint: disable=global-statement with gr.Accordion(open=False, label="Mask", elem_id="control_mask", elem_classes=["small-accordion"]): controls.clear() with gr.Row(): - controls.append(gr.Checkbox(label="Live update", value=True, elem_id="control_mask_live_update")) - btn_mask = ui_components.ToolButton(value=ui_symbols.refresh, visible=True, elem_id="control_mask_refresh", ) - btn_lama = ui_components.ToolButton(value=ui_symbols.image, visible=True, elem_id="control_mask_remove") + controls.append(gr.Checkbox(label="Live update", value=False, visible=False, elem_id="control_mask_live_update")) with gr.Row(): controls.append(gr.Checkbox(label="Inpaint masked only", value=False, elem_id="control_mask_only", )) controls.append(gr.Checkbox(label="Invert mask", value=False, elem_id="control_mask_invert")) with gr.Row(): - controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Blur', value=0.01, elem_id="control_mask_blur")) - controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Erode', value=0.01, elem_id="control_mask_erode")) - controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Dilate', value=0.01, elem_id="control_mask_dilate")) + controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Dilate', value=0, elem_id="control_mask_dilate")) + controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Erode', value=0, elem_id="control_mask_erode")) + with gr.Row(): + controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Blur', value=0, elem_id="control_mask_blur")) + controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Auto min score', value=0.8, elem_id="control_mask_score")) with gr.Row(): controls.append(gr.Dropdown(label="Auto-mask", choices=['None', 'Threshold', 'Edge', 'Grayscale'], value='None', elem_id="control_mask_auto")) - selected_model = gr.Dropdown(label="Auto-segment", choices=MODELS.keys(), value='None', elem_id="control_mask_segment") + controls.append(gr.Dropdown(label="Auto-segment", choices=MODELS.keys(), value='None', elem_id="control_mask_segment")) with gr.Row(): - controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='Score', value=0.5, visible=False, elem_id="control_mask_score")) controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='IOU', value=0.5, visible=False, elem_id="control_mask_iou")) controls.append(gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='NMS', value=0.5, visible=False, elem_id="control_mask_nms")) with gr.Row(): controls.append(gr.Dropdown(label="Preview", choices=['None', 'Masked', 'Binary', 'Grayscale', 'Color', 'Composite'], value='Composite', elem_id="control_mask_preview")) controls.append(gr.Dropdown(label="Colormap", choices=COLORMAP, value='pink', elem_id="control_mask_colormap")) + with gr.Row(): + btn_mask = gr.Button("Run Preview", elem_id="control_mask_refresh", ) + btn_lama = gr.Button("LaMa Remove", elem_id="control_mask_remove") - selected_model.change(fn=init_model, inputs=[selected_model], outputs=[selected_model]) for control in controls: control.change(fn=update_opts, inputs=controls, outputs=[]) return controls @@ -566,3 +568,28 @@ def bind_controls(image_controls: List[gr.Image], preview_image: gr.Image, outpu image_control.edit(fn=run_mask_live, inputs=[image_control], outputs=[preview_image]) for control in controls: control.change(fn=run_mask_live, inputs=[image_control], outputs=[preview_image]) + + +def process_kanvas(kanvas_data): + from modules import ui_control_helpers + if kanvas_data is None or 'kanvas' not in kanvas_data: + return None + input_image, input_mask = ui_control_helpers.process_kanvas(kanvas_data) + shared.log.debug(f'Kanvas mask: opts={vars(opts)}') + output_mask = run_mask(input_image, input_mask) + return output_mask + + +def process_kanvas_lama(kanvas_data): + from modules import ui_control_helpers + if kanvas_data is None or 'kanvas' not in kanvas_data: + return None + input_image, input_mask = ui_control_helpers.process_kanvas(kanvas_data) + shared.log.debug(f'Kanvas LaMa: opts={vars(opts)}') + output_mask = run_lama(input_image, input_mask) + return output_mask + + +def bind_kanvas(input_image: Image.Image, output_image: gr.Image): + btn_mask.click(_js='getKanvasData', fn=process_kanvas, inputs=[input_image], outputs=[output_image]) + btn_lama.click(_js='getKanvasData', fn=process_kanvas_lama, inputs=[input_image], outputs=[output_image]) diff --git a/modules/processing_class.py b/modules/processing_class.py index 660000c7c..836af33b8 100644 --- a/modules/processing_class.py +++ b/modules/processing_class.py @@ -500,28 +500,27 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): self.ops.append('img2img') crop_region = None - if self.image_mask is not None: - if type(self.image_mask) == list: - self.image_mask = self.image_mask[0] - if 'control' in self.ops: - self.image_mask = masking.run_mask(input_image=self.init_images, input_mask=self.image_mask, return_type='Grayscale', invert=self.inpainting_mask_invert==1) # blur/padding are handled in masking module - else: - self.image_mask = masking.run_mask(input_image=self.init_images, input_mask=self.image_mask, return_type='Grayscale', invert=self.inpainting_mask_invert==1, mask_blur=self.mask_blur, mask_padding=self.inpaint_full_res_padding) # old img2img - if self.inpaint_full_res: # mask only inpaint - self.mask_for_overlay = self.image_mask - mask = self.image_mask.convert('L') - crop_region = masking.get_crop_region(np.array(mask), self.inpaint_full_res_padding) - crop_region = masking.expand_crop_region(crop_region, self.width, self.height, mask.width, mask.height) - x1, y1, x2, y2 = crop_region - crop_mask = mask.crop(crop_region) - self.image_mask = images.resize_image(resize_mode=2, im=crop_mask, width=self.width, height=self.height) - self.paste_to = (x1, y1, x2-x1, y2-y1) - else: # full image inpaint - self.image_mask = images.resize_image(resize_mode=self.resize_mode, im=self.image_mask, width=self.width, height=self.height) - np_mask = np.array(self.image_mask) - np_mask = np.clip((np_mask.astype(np.float32)) * 2, 0, 255).astype(np.uint8) - self.mask_for_overlay = Image.fromarray(np_mask) - self.overlay_images = [] + if type(self.image_mask) == list: + self.image_mask = self.image_mask[0] + if 'Control' in self.__class__.__name__: + self.image_mask = masking.run_mask(input_image=self.init_images, input_mask=self.image_mask, return_type='Grayscale', invert=self.inpainting_mask_invert==1) # blur/padding are handled in masking module + elif self.image_mask is not None: + self.image_mask = masking.run_mask(input_image=self.init_images, input_mask=self.image_mask, return_type='Grayscale', invert=self.inpainting_mask_invert==1, mask_blur=self.mask_blur, mask_padding=self.inpaint_full_res_padding) # old img2img + if self.inpaint_full_res and self.image_mask is not None: # mask only inpaint + self.mask_for_overlay = self.image_mask + mask = self.image_mask.convert('L') + crop_region = masking.get_crop_region(np.array(mask), self.inpaint_full_res_padding) + crop_region = masking.expand_crop_region(crop_region, self.width, self.height, mask.width, mask.height) + x1, y1, x2, y2 = crop_region + crop_mask = mask.crop(crop_region) + self.image_mask = images.resize_image(resize_mode=2, im=crop_mask, width=self.width, height=self.height) + self.paste_to = (x1, y1, x2-x1, y2-y1) + elif self.image_mask is not None: # full image inpaint + self.image_mask = images.resize_image(resize_mode=self.resize_mode, im=self.image_mask, width=self.width, height=self.height) + np_mask = np.array(self.image_mask) + np_mask = np.clip((np_mask.astype(np.float32)) * 2, 0, 255).astype(np.uint8) + self.mask_for_overlay = Image.fromarray(np_mask) + self.overlay_images = [] add_color_corrections = shared.opts.img2img_color_correction and self.color_corrections is None if add_color_corrections: @@ -563,7 +562,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing): processed_images.append(image) self.init_images = processed_images # self.batch_size = len(self.init_images) - if self.overlay_images is not None: + if self.overlay_images is not None and len(self.overlay_images) > 0: self.overlay_images = self.overlay_images * self.batch_size if self.color_corrections is not None and len(self.color_corrections) == 1: self.color_corrections = self.color_corrections * self.batch_size diff --git a/modules/ui_control.py b/modules/ui_control.py index 9b9f0f287..12799a5df 100644 --- a/modules/ui_control.py +++ b/modules/ui_control.py @@ -262,7 +262,6 @@ def create_ui(_blocks: gr.Blocks=None): btn_interrogate.click(**select_dict) # need to fetch input first btn_interrogate.click(fn=helpers.interrogate, inputs=[], outputs=[prompt]) - prompt.submit(**select_dict) negative.submit(**select_dict) btn_generate.click(**select_dict) @@ -412,7 +411,11 @@ def create_ui(_blocks: gr.Blocks=None): generation_parameters_copypaste.add_paste_fields("control", input_image, paste_fields, override_settings) bindings = generation_parameters_copypaste.ParamBinding(paste_button=btn_paste, tabname="control", source_text_component=prompt, source_image_component=output_gallery) generation_parameters_copypaste.register_paste_params_button(bindings) - # masking.bind_controls([input_image], preview_process, output_image) + + if (installer.version['kanvas'] == 'disabled') or (installer.version['kanvas'] == 'unavailable'): + masking.bind_controls([input_image], preview_process, output_image) + else: + masking.bind_kanvas(input_image, preview_process) if os.environ.get('SD_CONTROL_DEBUG', None) is not None: # debug only from modules.control.test import test_processors, test_controlnets, test_adapters, test_xs, test_lite