add sd3.5

Signed-off-by: Vladimir Mandic <mandic00@live.com>

CHANGELOG.md

@@ -19,6 +19,7 @@
 
 - New fine-tuned [CLiP-ViT-L]((https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14)) 1st stage **text-encoders** used by SD15, SDXL, Flux.1, etc. brings additional details to your images  
 - New models:  
+  [Stable Diffusion 3.5 Large](https://huggingface.co/stabilityai/stable-diffusion-3.5-large)  
   [OmniGen](https://arxiv.org/pdf/2409.11340)  
   [CogView 3 Plus](https://huggingface.co/THUDM/CogView3-Plus-3B)  
   [Meissonic](https://github.com/viiika/Meissonic)  
@@ -178,6 +179,12 @@ And there are also other goodies like multiple *XYZ grid* improvements, addition
   - Params used: prompt, steps, guidance scale for prompt guidance, refine guidance scale for image guidance  
     Recommended: guidance=3.0, refine-guidance=1.6  
 
+- [Stable Diffusion 3.5 Large](https://huggingface.co/stabilityai/stable-diffusion-3.5-large)  
+  - New/improved variant of Stable Diffusion 3  
+  - Select from *networks -> models -> reference*  
+  - Available in standard and turbo variations  
+  - *Note*: Access to to both variations of SD3.5 model is gated, you must accept the conditions and use HF login  
+
 - [CogView 3 Plus](https://huggingface.co/THUDM/CogView3-Plus-3B)
   - Select from *networks -> models -> reference*  
   - resolution width and height can be from 512px to 2048px and must be divisible by 32  

html/reference.json

@@ -112,21 +112,27 @@
     "extras": "sampler: Default, cfg_scale: 4.0, image_cfg_scale: 1.0"
   },
   "StabilityAI Stable Diffusion 3 Medium": {
-    "path": "huggingface/stabilityai/stable-diffusion-3-medium-diffusers",
+    "path": "stabilityai/stable-diffusion-3-medium-diffusers",
     "skip": true,
     "variant": "fp16",
-    "te3": null,
     "desc": "Stable Diffusion 3 Medium is a Multimodal Diffusion Transformer (MMDiT) text-to-image model that features greatly improved performance in image quality, typography, complex prompt understanding, and resource-efficiency",
     "preview": "stabilityai--stable-diffusion-3.jpg",
     "extras": "sampler: Default, cfg_scale: 7.0"
   },
-  "StabilityAI Stable Diffusion 3 Large": {
-    "path": "huggingface/stabilityai/stable-diffusion-3.5-large",
+  "StabilityAI Stable Diffusion 3.5 Large": {
+    "path": "stabilityai/stable-diffusion-3.5-large",
     "skip": true,
     "variant": "fp16",
-    "te3": null,
     "desc": "Stable Diffusion 3 Medium is a Multimodal Diffusion Transformer (MMDiT) text-to-image model that features greatly improved performance in image quality, typography, complex prompt understanding, and resource-efficiency",
-    "preview": "stabilityai--stable-diffusion-3.jpg",
+    "preview": "stabilityai--stable-diffusion-3_5.jpg",
+    "extras": "sampler: Default, cfg_scale: 7.0"
+  },
+  "StabilityAI Stable Diffusion 3.5 Turbo": {
+    "path": "stabilityai/stable-diffusion-3.5-large-turbo",
+    "skip": true,
+    "variant": "fp16",
+    "desc": "Stable Diffusion 3 Medium is a Multimodal Diffusion Transformer (MMDiT) text-to-image model that features greatly improved performance in image quality, typography, complex prompt understanding, and resource-efficiency",
+    "preview": "stabilityai--stable-diffusion-3_5.jpg",
     "extras": "sampler: Default, cfg_scale: 7.0"
   },
 

modules/model_sd3.py

@@ -1,5 +1,4 @@
 import os
-import torch
 import diffusers
 import transformers
 
@@ -7,7 +6,6 @@ import transformers
 def load_sd3(checkpoint_info, cache_dir=None, config=None):
     from modules import devices, modelloader, sd_models
     repo_id = sd_models.path_to_repo(checkpoint_info.name)
-    # dtype = torch.float16
     dtype = devices.dtype
     kwargs = {}
     if checkpoint_info.path is not None and checkpoint_info.path.endswith('.safetensors') and os.path.exists(checkpoint_info.path):
@@ -48,6 +46,7 @@ def load_sd3(checkpoint_info, cache_dir=None, config=None):
     else:
         modelloader.hf_login()
         loader = diffusers.StableDiffusion3Pipeline.from_pretrained
+        kwargs['variant'] = 'fp16'
     pipe = loader(
         repo_id,
         torch_dtype=dtype,
@@ -55,6 +54,5 @@ def load_sd3(checkpoint_info, cache_dir=None, config=None):
         config=config,
         **kwargs,
     )
-    # pipe.transformer = pipe.transformer.to(devices.dtype) # diffusers loader leaves it as-is
     devices.torch_gc()
     return pipe

modules/modelloader.py

@@ -196,6 +196,8 @@ def download_diffusers_model(hub_id: str, cache_dir: str = None, download_config
         return None
     from diffusers import DiffusionPipeline
     shared.state.begin('HuggingFace')
+    if hub_id.startswith('huggingface/'):
+        hub_id = hub_id.replace('huggingface/', '')
     if download_config is None:
         download_config = {
             "force_download": False,
@@ -335,7 +337,7 @@ def get_reference_opts(name: str, quiet=False):
         # shared.log.error(f'Reference: model="{name}" not found')
         return {}
     if not quiet:
-        shared.log.debug(f'Reference: model="{name}" {model_opts.get("extras", None)}')
+        shared.log.debug(f'Reference: model="{name}" {model_opts}')
     return model_opts
 
 

modules/prompt_parser_diffusers.py

@@ -127,7 +127,11 @@ def get_tokens(msg, prompt):
     if shared.sd_loaded and hasattr(shared.sd_model, 'tokenizer') and shared.sd_model.tokenizer is not None:
         if token_dict is None or token_type != shared.sd_model_type:
             token_type = shared.sd_model_type
-            fn = os.path.join(shared.sd_model.tokenizer.name_or_path, 'tokenizer', 'vocab.json')
+            fn = shared.sd_model.tokenizer.name_or_path
+            if fn.endswith('tokenizer'):
+                fn = os.path.join(shared.sd_model.tokenizer.name_or_path, 'vocab.json')
+            else:
+                fn = os.path.join(shared.sd_model.tokenizer.name_or_path, 'tokenizer', 'vocab.json')
             token_dict = shared.readfile(fn, silent=True)
             for k, v in shared.sd_model.tokenizer.added_tokens_decoder.items():
                 token_dict[str(v)] = k