Revert "fix"

This reverts commit f90a5139a2.

.github/workflows/pr_test_fetcher.yml

@@ -1,6 +1,12 @@
 name: Fast tests for PRs - Test Fetcher
 
-on: workflow_dispatch
+on:
+  pull_request:
+    branches:
+      - main
+  push:
+    branches:
+      - ci-*
 
 env:
   DIFFUSERS_IS_CI: yes

examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py

@@ -133,7 +133,7 @@ def save_model_card(
         diffusers_imports_pivotal = """from huggingface_hub import hf_hub_download
 from safetensors.torch import load_file
         """
-        diffusers_example_pivotal = f"""embedding_path = hf_hub_download(repo_id='{repo_id}', filename="embeddings.safetensors", repo_type="model")
+        diffusers_example_pivotal = f"""embedding_path = hf_hub_download(repo_id="{repo_id}", filename="embeddings.safetensors", repo_type="model")
 state_dict = load_file(embedding_path)
 pipeline.load_textual_inversion(state_dict["clip_l"], token=["<s0>", "<s1>"], text_encoder=pipe.text_encoder, tokenizer=pipe.tokenizer)
 pipeline.load_textual_inversion(state_dict["clip_g"], token=["<s0>", "<s1>"], text_encoder=pipe.text_encoder_2, tokenizer=pipe.tokenizer_2)
@@ -145,7 +145,8 @@ pipeline.load_textual_inversion(state_dict["clip_g"], token=["<s0>", "<s1>"], te
 to trigger concept `{key}` → use `{tokens}` in your prompt \n
 """
 
-    yaml = f"""---
+    yaml = f"""
+---
 tags:
 - stable-diffusion-xl
 - stable-diffusion-xl-diffusers
@@ -158,7 +159,7 @@ base_model: {base_model}
 instance_prompt: {instance_prompt}
 license: openrail++
 ---
-"""
+    """
 
     model_card = f"""
 # SDXL LoRA DreamBooth - {repo_id}
@@ -169,6 +170,14 @@ license: openrail++
 
 ### These are {repo_id} LoRA adaption weights for {base_model}.
 
+The weights were trained  using [DreamBooth](https://dreambooth.github.io/).
+
+LoRA for the text encoder was enabled: {train_text_encoder}.
+
+Pivotal tuning was enabled: {train_text_encoder_ti}.
+
+Special VAE used for training: {vae_path}.
+
 ## Trigger words
 
 {trigger_str}
@@ -187,24 +196,11 @@ image = pipeline('{validation_prompt if validation_prompt else instance_prompt}'
 
 For more details, including weighting, merging and fusing LoRAs, check the [documentation on loading LoRAs in diffusers](https://huggingface.co/docs/diffusers/main/en/using-diffusers/loading_adapters)
 
-## Download model
+## Download model (use it with UIs such as AUTO1111, Comfy, SD.Next, Invoke)
 
-### Use it with UIs such as AUTOMATIC1111, Comfy UI, SD.Next, Invoke
+Weights for this model are available in Safetensors format.
 
-- Download the LoRA *.safetensors [here](/{repo_id}/blob/main/pytorch_lora_weights.safetensors). Rename it and place it on your Lora folder.
-- Download the text embeddings *.safetensors [here](/{repo_id}/blob/main/embeddings.safetensors). Rename it and place it on it on your embeddings folder.
-
-All [Files & versions](/{repo_id}/tree/main).
-
-## Details
-
-The weights were trained using [🧨 diffusers Advanced Dreambooth Training Script](https://github.com/huggingface/diffusers/blob/main/examples/advanced_diffusion_training/train_dreambooth_lora_sdxl_advanced.py).
-
-LoRA for the text encoder was enabled. {train_text_encoder}.
-
-Pivotal tuning was enabled: {train_text_encoder_ti}.
-
-Special VAE used for training: {vae_path}.
+[Download]({repo_id}/tree/main) them in the Files & versions tab.
 
 """
     with open(os.path.join(repo_folder, "README.md"), "w") as f:
@@ -671,12 +667,6 @@ def parse_args(input_args=None):
         default=4,
         help=("The dimension of the LoRA update matrices."),
     )
-    parser.add_argument(
-        "--cache_latents",
-        action="store_true",
-        default=False,
-        help="Cache the VAE latents",
-    )
 
     if input_args is not None:
         args = parser.parse_args(input_args)
@@ -1180,7 +1170,6 @@ def main(args):
         revision=args.revision,
         variant=args.variant,
     )
-    vae_scaling_factor = vae.config.scaling_factor
     unet = UNet2DConditionModel.from_pretrained(
         args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
     )
@@ -1611,20 +1600,6 @@ def main(args):
             args.validation_prompt = args.validation_prompt.replace(token_abs, "".join(token_replacement))
     print("validation prompt:", args.validation_prompt)
 
-    if args.cache_latents:
-        latents_cache = []
-        for batch in tqdm(train_dataloader, desc="Caching latents"):
-            with torch.no_grad():
-                batch["pixel_values"] = batch["pixel_values"].to(
-                    accelerator.device, non_blocking=True, dtype=torch.float32
-                )
-                latents_cache.append(vae.encode(batch["pixel_values"]).latent_dist)
-
-        if args.validation_prompt is None:
-            del vae
-            if torch.cuda.is_available():
-                torch.cuda.empty_cache()
-
     # Scheduler and math around the number of training steps.
     overrode_max_train_steps = False
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
@@ -1740,7 +1715,9 @@ def main(args):
         unet.train()
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
+                pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
                 prompts = batch["prompts"]
+                # print(prompts)
                 # encode batch prompts when custom prompts are provided for each image -
                 if train_dataset.custom_instance_prompts:
                     if freeze_text_encoder:
@@ -1752,13 +1729,9 @@ def main(args):
                         tokens_one = tokenize_prompt(tokenizer_one, prompts, add_special_tokens)
                         tokens_two = tokenize_prompt(tokenizer_two, prompts, add_special_tokens)
 
-                if args.cache_latents:
-                    model_input = latents_cache[step].sample()
-                else:
-                    pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
-                    model_input = vae.encode(pixel_values).latent_dist.sample()
-
-                model_input = model_input * vae_scaling_factor
+                # Convert images to latent space
+                model_input = vae.encode(pixel_values).latent_dist.sample()
+                model_input = model_input * vae.config.scaling_factor
                 if args.pretrained_vae_model_name_or_path is None:
                     model_input = model_input.to(weight_dtype)
 

examples/consistency_distillation/README.md

@@ -1,6 +1,6 @@
 # Latent Consistency Distillation Example:
 
-[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is a method to distill a latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use latent consistency distillation to distill stable-diffusion-v1.5 for inference with few timesteps.
+[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is method to distill latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use the latent consistency distillation to distill stable-diffusion-v1.5 for less timestep inference.
 
 ## Full model distillation
 
@@ -24,7 +24,7 @@ Then cd in the example folder and run
 pip install -r requirements.txt
 ```
 
-And initialize an [🤗 Accelerate](https://github.com/huggingface/accelerate/) environment with:
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
 
 ```bash
 accelerate config
@@ -46,16 +46,12 @@ write_basic_config()
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
 
 
-#### Example
-
-The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example, and for illustrative purposes only. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/). You may also need to search the hyperparameter space according to the dataset you use.
+#### Example with LAION-A6+ dataset
 
 ```bash
-export MODEL_NAME="runwayml/stable-diffusion-v1-5"
-export OUTPUT_DIR="path/to/saved/model"
-
-accelerate launch train_lcm_distill_sd_wds.py \
-    --pretrained_teacher_model=$MODEL_NAME \
+runwayml/stable-diffusion-v1-5
+PROGRAM="train_lcm_distill_sd_wds.py \
+    --pretrained_teacher_model=$MODEL_DIR \
     --output_dir=$OUTPUT_DIR \
     --mixed_precision=fp16 \
     --resolution=512 \
@@ -63,7 +59,7 @@ accelerate launch train_lcm_distill_sd_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
+    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \
@@ -73,23 +69,19 @@ accelerate launch train_lcm_distill_sd_wds.py \
     --resume_from_checkpoint=latest \
     --report_to=wandb \
     --seed=453645634 \
-    --push_to_hub
+    --push_to_hub \
 ```
 
 ## LCM-LoRA
 
 Instead of fine-tuning the full model, we can also just train a LoRA that can be injected into any SDXL model.
 
-### Example
-
-The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/).
-
+### Example with LAION-A6+ dataset
+    
 ```bash
-export MODEL_NAME="runwayml/stable-diffusion-v1-5"
-export OUTPUT_DIR="path/to/saved/model"
-
-accelerate launch train_lcm_distill_lora_sd_wds.py \
-    --pretrained_teacher_model=$MODEL_NAME \
+runwayml/stable-diffusion-v1-5
+PROGRAM="train_lcm_distill_lora_sd_wds.py \
+    --pretrained_teacher_model=$MODEL_DIR \
     --output_dir=$OUTPUT_DIR \
     --mixed_precision=fp16 \
     --resolution=512 \
@@ -98,7 +90,7 @@ accelerate launch train_lcm_distill_lora_sd_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
+    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \

examples/consistency_distillation/README_sdxl.md

@@ -1,6 +1,6 @@
 # Latent Consistency Distillation Example:
 
-[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is a method to distill a latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use latent consistency distillation to distill SDXL for inference with few timesteps.
+[Latent Consistency Models (LCMs)](https://arxiv.org/abs/2310.04378) is method to distill latent diffusion model to enable swift inference with minimal steps. This example demonstrates how to use the latent consistency distillation to distill SDXL for less timestep inference.
 
 ## Full model distillation
 
@@ -24,7 +24,7 @@ Then cd in the example folder and run
 pip install -r requirements.txt
 ```
 
-And initialize an [🤗 Accelerate](https://github.com/huggingface/accelerate/) environment with:
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
 
 ```bash
 accelerate config
@@ -46,16 +46,12 @@ write_basic_config()
 When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
 
 
-#### Example
-
-The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example, and for illustrative purposes only. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/). You may also need to search the hyperparameter space according to the dataset you use.
+#### Example with LAION-A6+ dataset
 
 ```bash
-export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0"
-export OUTPUT_DIR="path/to/saved/model"
-
-accelerate launch train_lcm_distill_sdxl_wds.py \
-    --pretrained_teacher_model=$MODEL_NAME \
+export MODEL_DIR="stabilityai/stable-diffusion-xl-base-1.0"
+PROGRAM="train_lcm_distill_sdxl_wds.py \
+    --pretrained_teacher_model=$MODEL_DIR \
     --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
     --output_dir=$OUTPUT_DIR \
     --mixed_precision=fp16 \
@@ -64,7 +60,7 @@ accelerate launch train_lcm_distill_sdxl_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
+    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \
@@ -81,15 +77,11 @@ accelerate launch train_lcm_distill_sdxl_wds.py \
 
 Instead of fine-tuning the full model, we can also just train a LoRA that can be injected into any SDXL model.
 
-### Example
-
-The following uses the [Conceptual Captions 12M (CC12M) dataset](https://github.com/google-research-datasets/conceptual-12m) as an example. For best results you may consider large and high-quality text-image datasets such as [LAION](https://laion.ai/blog/laion-400-open-dataset/).
-
+### Example with LAION-A6+ dataset
+    
 ```bash
-export MODEL_NAME="stabilityai/stable-diffusion-xl-base-1.0"
-export OUTPUT_DIR="path/to/saved/model"
-
-accelerate launch train_lcm_distill_lora_sdxl_wds.py \
+export MODEL_DIR="stabilityai/stable-diffusion-xl-base-1.0"
+PROGRAM="train_lcm_distill_lora_sdxl_wds.py \
     --pretrained_teacher_model=$MODEL_DIR \
     --pretrained_vae_model_name_or_path=madebyollin/sdxl-vae-fp16-fix \
     --output_dir=$OUTPUT_DIR \
@@ -100,7 +92,7 @@ accelerate launch train_lcm_distill_lora_sdxl_wds.py \
     --max_train_steps=1000 \
     --max_train_samples=4000000 \
     --dataloader_num_workers=8 \
-    --train_shards_path_or_url="pipe:curl -L -s https://huggingface.co/datasets/laion/conceptual-captions-12m-webdataset/resolve/main/data/{00000..01099}.tar?download=true" \
+    --train_shards_path_or_url='pipe:aws s3 cp s3://muse-datasets/laion-aesthetic6plus-min512-data/{00000..01210}.tar -' \
     --validation_steps=200 \
     --checkpointing_steps=200 --checkpoints_total_limit=10 \
     --train_batch_size=12 \

examples/consistency_distillation/train_lcm_distill_lora_sd_wds.py

@@ -1123,7 +1123,7 @@ def main(args):
     for epoch in range(first_epoch, args.num_train_epochs):
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
-                image, text = batch
+                image, text, _, _ = batch
 
                 image = image.to(accelerator.device, non_blocking=True)
                 encoded_text = compute_embeddings_fn(text)

examples/consistency_distillation/train_lcm_distill_lora_sdxl_wds.py

@@ -68,11 +68,6 @@ from diffusers.utils.import_utils import is_xformers_available
 
 MAX_SEQ_LENGTH = 77
 
-# Adjust for your dataset
-WDS_JSON_WIDTH = "width"  # original_width for LAION
-WDS_JSON_HEIGHT = "height"  # original_height for LAION
-MIN_SIZE = 700  # ~960 for LAION, ideal: 1024 if the dataset contains large images
-
 if is_wandb_available():
     import wandb
 
@@ -151,10 +146,10 @@ class WebdatasetFilter:
         try:
             if "json" in x:
                 x_json = json.loads(x["json"])
-                filter_size = (x_json.get(WDS_JSON_WIDTH, 0.0) or 0.0) >= self.min_size and x_json.get(
-                    WDS_JSON_HEIGHT, 0
+                filter_size = (x_json.get("original_width", 0.0) or 0.0) >= self.min_size and x_json.get(
+                    "original_height", 0
                 ) >= self.min_size
-                filter_watermark = (x_json.get("pwatermark", 0.0) or 0.0) <= self.max_pwatermark
+                filter_watermark = (x_json.get("pwatermark", 1.0) or 1.0) <= self.max_pwatermark
                 return filter_size and filter_watermark
             else:
                 return False
@@ -185,7 +180,7 @@ class Text2ImageDataset:
             if use_fix_crop_and_size:
                 return (resolution, resolution)
             else:
-                return (int(json.get(WDS_JSON_WIDTH, 0.0)), int(json.get(WDS_JSON_HEIGHT, 0.0)))
+                return (int(json.get("original_width", 0.0)), int(json.get("original_height", 0.0)))
 
         def transform(example):
             # resize image
@@ -217,7 +212,7 @@ class Text2ImageDataset:
         pipeline = [
             wds.ResampledShards(train_shards_path_or_url),
             tarfile_to_samples_nothrow,
-            wds.select(WebdatasetFilter(min_size=MIN_SIZE)),
+            wds.select(WebdatasetFilter(min_size=960)),
             wds.shuffle(shuffle_buffer_size),
             *processing_pipeline,
             wds.batched(per_gpu_batch_size, partial=False, collation_fn=default_collate),

examples/consistency_distillation/train_lcm_distill_sd_wds.py

@@ -1106,7 +1106,7 @@ def main(args):
     for epoch in range(first_epoch, args.num_train_epochs):
         for step, batch in enumerate(train_dataloader):
             with accelerator.accumulate(unet):
-                image, text = batch
+                image, text, _, _ = batch
 
                 image = image.to(accelerator.device, non_blocking=True)
                 encoded_text = compute_embeddings_fn(text)

examples/consistency_distillation/train_lcm_distill_sdxl_wds.py

@@ -67,11 +67,6 @@ from diffusers.utils.import_utils import is_xformers_available
 
 MAX_SEQ_LENGTH = 77
 
-# Adjust for your dataset
-WDS_JSON_WIDTH = "width"  # original_width for LAION
-WDS_JSON_HEIGHT = "height"  # original_height for LAION
-MIN_SIZE = 700  # ~960 for LAION, ideal: 1024 if the dataset contains large images
-
 if is_wandb_available():
     import wandb
 
@@ -133,10 +128,10 @@ class WebdatasetFilter:
         try:
             if "json" in x:
                 x_json = json.loads(x["json"])
-                filter_size = (x_json.get(WDS_JSON_WIDTH, 0.0) or 0.0) >= self.min_size and x_json.get(
-                    WDS_JSON_HEIGHT, 0
+                filter_size = (x_json.get("original_width", 0.0) or 0.0) >= self.min_size and x_json.get(
+                    "original_height", 0
                 ) >= self.min_size
-                filter_watermark = (x_json.get("pwatermark", 0.0) or 0.0) <= self.max_pwatermark
+                filter_watermark = (x_json.get("pwatermark", 1.0) or 1.0) <= self.max_pwatermark
                 return filter_size and filter_watermark
             else:
                 return False
@@ -167,7 +162,7 @@ class Text2ImageDataset:
             if use_fix_crop_and_size:
                 return (resolution, resolution)
             else:
-                return (int(json.get(WDS_JSON_WIDTH, 0.0)), int(json.get(WDS_JSON_HEIGHT, 0.0)))
+                return (int(json.get("original_width", 0.0)), int(json.get("original_height", 0.0)))
 
         def transform(example):
             # resize image
@@ -199,7 +194,7 @@ class Text2ImageDataset:
         pipeline = [
             wds.ResampledShards(train_shards_path_or_url),
             tarfile_to_samples_nothrow,
-            wds.select(WebdatasetFilter(min_size=MIN_SIZE)),
+            wds.select(WebdatasetFilter(min_size=960)),
             wds.shuffle(shuffle_buffer_size),
             *processing_pipeline,
             wds.batched(per_gpu_batch_size, partial=False, collation_fn=default_collate),

src/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py

@@ -446,9 +446,8 @@ def convert_ldm_unet_checkpoint(
         new_checkpoint["add_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"]
 
     # Relevant to StableDiffusionUpscalePipeline
-    if "num_class_embeds" in config:
-        if (config["num_class_embeds"] is not None) and ("label_emb.weight" in unet_state_dict):
-            new_checkpoint["class_embedding.weight"] = unet_state_dict["label_emb.weight"]
+    if (config["num_class_embeds"] is not None) and ("label_emb.weight" in unet_state_dict):
+        new_checkpoint["class_embedding.weight"] = unet_state_dict["label_emb.weight"]
 
     new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
     new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]