fix hires batch, sdxl refiner

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

CHANGELOG.md

@@ -20,6 +20,8 @@
   - interrogate caption with T5  
   - on-the-fly quantization using TorchAO  
   - remove concurrent preview requests  
+  - hires batch  
+  - sdxl refiner  
 
 ## Update for 2024-12-24
 

configs/sdxl-refiner/model_index.json

@@ -0,0 +1,35 @@
+{
+  "_class_name": "StableDiffusionXLImg2ImgPipeline",
+  "_diffusers_version": "0.19.0.dev0",
+  "force_zeros_for_empty_prompt": false,
+  "add_watermarker": null,
+  "requires_aesthetics_score": true,
+  "scheduler": [
+    "diffusers",
+    "EulerDiscreteScheduler"
+  ],
+  "text_encoder": [
+    null,
+    null
+  ],
+  "text_encoder_2": [
+    "transformers",
+    "CLIPTextModelWithProjection"
+  ],
+  "tokenizer": [
+    null,
+    null
+  ],
+  "tokenizer_2": [
+    "transformers",
+    "CLIPTokenizer"
+  ],
+  "unet": [
+    "diffusers",
+    "UNet2DConditionModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderKL"
+  ]
+}

configs/sdxl-refiner/scheduler/scheduler_config.json

@@ -0,0 +1,18 @@
+{
+  "_class_name": "EulerDiscreteScheduler",
+  "_diffusers_version": "0.19.0.dev0",
+  "beta_end": 0.012,
+  "beta_schedule": "scaled_linear",
+  "beta_start": 0.00085,
+  "clip_sample": false,
+  "interpolation_type": "linear",
+  "num_train_timesteps": 1000,
+  "prediction_type": "epsilon",
+  "sample_max_value": 1.0,
+  "set_alpha_to_one": false,
+  "skip_prk_steps": true,
+  "steps_offset": 1,
+  "timestep_spacing": "leading",
+  "trained_betas": null,
+  "use_karras_sigmas": false
+}

configs/sdxl-refiner/text_encoder_2/config.json

@@ -0,0 +1,24 @@
+{
+  "architectures": [
+    "CLIPTextModelWithProjection"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 20,
+  "num_hidden_layers": 32,
+  "pad_token_id": 1,
+  "projection_dim": 1280,
+  "torch_dtype": "float16",
+  "transformers_version": "4.32.0.dev0",
+  "vocab_size": 49408
+}

configs/sdxl-refiner/tokenizer_2/special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "!",
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

configs/sdxl-refiner/tokenizer_2/tokenizer_config.json

@@ -0,0 +1,33 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "clean_up_tokenization_spaces": true,
+  "do_lower_case": true,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "model_max_length": 77,
+  "pad_token": "!",
+  "tokenizer_class": "CLIPTokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

configs/sdxl-refiner/unet/config.json

@@ -0,0 +1,69 @@
+{
+  "_class_name": "UNet2DConditionModel",
+  "_diffusers_version": "0.19.0.dev0",
+  "act_fn": "silu",
+  "addition_embed_type": "text_time",
+  "addition_embed_type_num_heads": 64,
+  "addition_time_embed_dim": 256,
+  "attention_head_dim": [
+    6,
+    12,
+    24,
+    24
+  ],
+  "block_out_channels": [
+    384,
+    768,
+    1536,
+    1536
+  ],
+  "center_input_sample": false,
+  "class_embed_type": null,
+  "class_embeddings_concat": false,
+  "conv_in_kernel": 3,
+  "conv_out_kernel": 3,
+  "cross_attention_dim": 1280,
+  "cross_attention_norm": null,
+  "down_block_types": [
+    "DownBlock2D",
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D",
+    "DownBlock2D"
+  ],
+  "downsample_padding": 1,
+  "dual_cross_attention": false,
+  "encoder_hid_dim": null,
+  "encoder_hid_dim_type": null,
+  "flip_sin_to_cos": true,
+  "freq_shift": 0,
+  "in_channels": 4,
+  "layers_per_block": 2,
+  "mid_block_only_cross_attention": null,
+  "mid_block_scale_factor": 1,
+  "mid_block_type": "UNetMidBlock2DCrossAttn",
+  "norm_eps": 1e-05,
+  "norm_num_groups": 32,
+  "num_attention_heads": null,
+  "num_class_embeds": null,
+  "only_cross_attention": false,
+  "out_channels": 4,
+  "projection_class_embeddings_input_dim": 2560,
+  "resnet_out_scale_factor": 1.0,
+  "resnet_skip_time_act": false,
+  "resnet_time_scale_shift": "default",
+  "sample_size": 128,
+  "time_cond_proj_dim": null,
+  "time_embedding_act_fn": null,
+  "time_embedding_dim": null,
+  "time_embedding_type": "positional",
+  "timestep_post_act": null,
+  "transformer_layers_per_block": 4,
+  "up_block_types": [
+    "UpBlock2D",
+    "CrossAttnUpBlock2D",
+    "CrossAttnUpBlock2D",
+    "UpBlock2D"
+  ],
+  "upcast_attention": null,
+  "use_linear_projection": true
+}

configs/sdxl-refiner/vae/config.json

@@ -0,0 +1,32 @@
+{
+  "_class_name": "AutoencoderKL",
+  "_diffusers_version": "0.20.0.dev0",
+  "_name_or_path": "../sdxl-vae/",
+  "act_fn": "silu",
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": true,
+  "in_channels": 3,
+  "latent_channels": 4,
+  "layers_per_block": 2,
+  "norm_num_groups": 32,
+  "out_channels": 3,
+  "sample_size": 1024,
+  "scaling_factor": 0.13025,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ]
+}

modules/processing_args.py

@@ -6,6 +6,7 @@ import time
 import inspect
 import torch
 import numpy as np
+from PIL import Image
 from modules import shared, errors, sd_models, processing, processing_vae, processing_helpers, sd_hijack_hypertile, prompt_parser_diffusers, timer, extra_networks
 from modules.processing_callbacks import diffusers_callback_legacy, diffusers_callback, set_callbacks_p
 from modules.processing_helpers import resize_hires, fix_prompts, calculate_base_steps, calculate_hires_steps, calculate_refiner_steps, get_generator, set_latents, apply_circular # pylint: disable=unused-import
@@ -22,7 +23,8 @@ def task_specific_kwargs(p, model):
     if len(getattr(p, 'init_images', [])) > 0:
         if isinstance(p.init_images[0], str):
             p.init_images = [helpers.decode_base64_to_image(i, quiet=True) for i in p.init_images]
-        p.init_images = [i.convert('RGB') if i.mode != 'RGB' else i for i in p.init_images if i is not None]
+        if isinstance(p.init_images[0], Image.Image):
+            p.init_images = [i.convert('RGB') if i.mode != 'RGB' else i for i in p.init_images if i is not None]
     if (sd_models.get_diffusers_task(model) == sd_models.DiffusersTaskType.TEXT_2_IMAGE or len(getattr(p, 'init_images', [])) == 0) and not is_img2img_model:
         p.ops.append('txt2img')
         if hasattr(p, 'width') and hasattr(p, 'height'):
@@ -99,7 +101,7 @@ def task_specific_kwargs(p, model):
     return task_args
 
 
-def set_pipeline_args(p, model, prompts: list, negative_prompts: list, prompts_2: typing.Optional[list]=None, negative_prompts_2: typing.Optional[list]=None, desc:str='', **kwargs):
+def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:typing.Optional[list]=None, negative_prompts_2:typing.Optional[list]=None, prompt_attention:typing.Optional[str]=None, desc:typing.Optional[str]='', **kwargs):
     t0 = time.time()
     shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model)
     apply_circular(p.tiling, model)
@@ -118,7 +120,8 @@ def set_pipeline_args(p, model, prompts: list, negative_prompts: list, prompts_2
     clip_skip = kwargs.pop("clip_skip", 1)
 
     parser = 'fixed'
-    if shared.opts.prompt_attention != 'fixed' and 'Onnx' not in model.__class__.__name__ and (
+    prompt_attention = prompt_attention or shared.opts.prompt_attention
+    if prompt_attention != 'fixed' and 'Onnx' not in model.__class__.__name__ and (
         'StableDiffusion' in model.__class__.__name__ or
         'StableCascade' in model.__class__.__name__ or
         'Flux' in model.__class__.__name__
@@ -265,7 +268,7 @@ def set_pipeline_args(p, model, prompts: list, negative_prompts: list, prompts_2
     elif 'callback' in possible:
         args['callback'] = diffusers_callback_legacy
 
-    if 'image' in kwargs and len(getattr(p, 'init_images', [])) == 0:
+    if 'image' in kwargs:
         p.init_images = kwargs['image'] if isinstance(kwargs['image'], list) else [kwargs['image']]
 
     # handle remaining args

modules/processing_diffusers.py

@@ -199,8 +199,6 @@ def process_hires(p: processing.StableDiffusionProcessing, output):
             if p.is_control and hasattr(p, 'task_args') and p.task_args.get('image', None) is not None:
                 if hasattr(shared.sd_model, "vae") and output.images is not None and len(output.images) > 0:
                     output.images = processing_vae.vae_decode(latents=output.images, model=shared.sd_model, full_quality=p.full_quality, output_type='pil', width=p.hr_upscale_to_x, height=p.hr_upscale_to_y) # controlnet cannnot deal with latent input
-                    p.init_images = output.images # replace so hires uses new output
-                    # p.task_args['image'] = output.images # replace so hires uses new output
             update_sampler(p, shared.sd_model, second_pass=True)
             orig_denoise = p.denoising_strength
             p.denoising_strength = strength
@@ -289,10 +287,8 @@ def process_refine(p: processing.StableDiffusionProcessing, output):
                 image = processing_vae.vae_decode(latents=image, model=shared.sd_model, full_quality=p.full_quality, output_type='pil', width=p.width, height=p.height)
                 p.extra_generation_params['Noise level'] = noise_level
                 output_type = 'np'
-            if p.task_args.get('image', None) is not None and output is not None: # replace input with output so it can be used by hires/refine
-                # p.task_args['image'] = image
-                p.init_images = [image]
             update_sampler(p, shared.sd_refiner, second_pass=True)
+            shared.opts.prompt_attention = 'fixed'
             refiner_args = set_pipeline_args(
                 p=p,
                 model=shared.sd_refiner,
@@ -309,6 +305,7 @@ def process_refine(p: processing.StableDiffusionProcessing, output):
                 image=image,
                 output_type=output_type,
                 clip_skip=p.clip_skip,
+                prompt_attention='fixed',
                 desc='Refiner',
             )
             shared.state.sampling_steps = refiner_args.get('prior_num_inference_steps', None) or p.steps or refiner_args.get('num_inference_steps', None)

modules/sd_detect.py

@@ -170,6 +170,8 @@ def get_load_config(model_file, model_type, config_type='yaml'):
             return 'configs/sd15'
         if model_type == 'Stable Diffusion XL':
             return 'configs/sdxl'
+        if model_type == 'Stable Diffusion XL Refiner':
+            return 'configs/sdxl-refiner'
         if model_type == 'Stable Diffusion 3':
             return 'configs/sd3'
         if model_type == 'FLUX':

modules/shared_items.py

@@ -68,7 +68,7 @@ def get_pipelines():
         'Stable Diffusion Instruct': getattr(diffusers, 'StableDiffusionInstructPix2PixPipeline', None),
         'Stable Diffusion Upscale': getattr(diffusers, 'StableDiffusionUpscalePipeline', None),
         'Stable Diffusion XL': getattr(diffusers, 'StableDiffusionXLPipeline', None),
-        'Stable Diffusion XL Refiner': getattr(diffusers, 'StableDiffusionXLPipeline', None),
+        'Stable Diffusion XL Refiner': getattr(diffusers, 'StableDiffusionXLImg2ImgPipeline', None),
         'Stable Diffusion XL Img2Img': getattr(diffusers, 'StableDiffusionXLImg2ImgPipeline', None),
         'Stable Diffusion XL Inpaint': getattr(diffusers, 'StableDiffusionXLInpaintPipeline', None),
         'Stable Diffusion XL Instruct': getattr(diffusers, 'StableDiffusionXLInstructPix2PixPipeline', None),

modules/textual_inversion/textual_inversion.py

@@ -418,7 +418,7 @@ class EmbeddingDatabase:
         self.word_embeddings.update(sorted_word_embeddings)
 
         displayed_embeddings = (tuple(self.word_embeddings.keys()), tuple(self.skipped_embeddings.keys()))
-        if self.previously_displayed_embeddings != displayed_embeddings:
+        if self.previously_displayed_embeddings != displayed_embeddings and shared.opts.diffusers_enable_embed:
             self.previously_displayed_embeddings = displayed_embeddings
             t1 = time.time()
             shared.log.info(f"Load network: type=embeddings loaded={len(self.word_embeddings)} skipped={len(self.skipped_embeddings)} time={t1-t0:.2f}")