update

* Updated lora inference instructions

* Update examples/dreambooth/README.md

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

* Update README.md

* Update README.md

---------

Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>

docker/diffusers-pytorch-compile-cuda/Dockerfile

@@ -26,9 +26,9 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3.9 -m pip install --no-cache-dir --upgrade pip && \
     python3.9 -m pip install --no-cache-dir \
-    torch==2.1.2 \
-    torchvision==0.16.2 \
-    torchaudio==2.1.2 \
+    torch \
+    torchvision \
+    torchaudio \
     invisible_watermark && \
     python3.9 -m pip install --no-cache-dir \
     accelerate \

docker/diffusers-pytorch-cpu/Dockerfile

@@ -25,9 +25,9 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3 -m pip install --no-cache-dir --upgrade pip && \
     python3 -m pip install --no-cache-dir \
-        torch==2.1.2 \
-        torchvision==0.16.2 \
-        torchaudio==2.1.2 \
+        torch \
+        torchvision \
+        torchaudio \
         invisible_watermark \
         --extra-index-url https://download.pytorch.org/whl/cpu && \
     python3 -m pip install --no-cache-dir \

docker/diffusers-pytorch-cuda/Dockerfile

@@ -25,9 +25,9 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3 -m pip install --no-cache-dir --upgrade pip && \
     python3 -m pip install --no-cache-dir \
-    torch==2.1.2 \
-    torchvision==0.16.2 \
-    torchaudio==2.1.2 \
+    torch \
+    torchvision \
+    torchaudio \
     invisible_watermark && \
     python3 -m pip install --no-cache-dir \
     accelerate \

docker/diffusers-pytorch-xformers-cuda/Dockerfile

@@ -25,9 +25,9 @@ ENV PATH="/opt/venv/bin:$PATH"
 # pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
 RUN python3 -m pip install --no-cache-dir --upgrade pip && \
     python3 -m pip install --no-cache-dir \
-        torch==2.1.2 \
-        torchvision==0.16.2 \
-        torchaudio==2.1.2 \
+        torch \
+        torchvision \
+        torchaudio \
         invisible_watermark && \
     python3 -m pip install --no-cache-dir \
         accelerate \

docs/source/en/using-diffusers/custom_pipeline_overview.md

@@ -56,6 +56,60 @@ pipeline = DiffusionPipeline.from_pretrained(
 )
 ```
 
+### Load from a local file
+
+Community pipelines can also be loaded from a local file if you pass a file path instead. The path to the passed directory must contain a `pipeline.py` file that contains the pipeline class in order to successfully load it.
+
+```py
+pipeline = DiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    custom_pipeline="./path/to/pipeline_directory/",
+    clip_model=clip_model,
+    feature_extractor=feature_extractor,
+    use_safetensors=True,
+)
+```
+
+### Load from a specific version
+
+By default, community pipelines are loaded from the latest stable version of Diffusers. To load a community pipeline from another version, use the `custom_revision` parameter.
+
+<hfoptions id="version">
+<hfoption id="main">
+
+For example, to load from the `main` branch:
+
+```py
+pipeline = DiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    custom_pipeline="clip_guided_stable_diffusion",
+    custom_revision="main",
+    clip_model=clip_model,
+    feature_extractor=feature_extractor,
+    use_safetensors=True,
+)
+```
+
+</hfoption>
+<hfoption id="older version">
+
+For example, to load from a previous version of Diffusers like `v0.25.0`:
+
+```py
+pipeline = DiffusionPipeline.from_pretrained(
+    "runwayml/stable-diffusion-v1-5",
+    custom_pipeline="clip_guided_stable_diffusion",
+    custom_revision="v0.25.0",
+    clip_model=clip_model,
+    feature_extractor=feature_extractor,
+    use_safetensors=True,
+)
+```
+
+</hfoption>
+</hfoptions>
+
+
 For more information about community pipelines, take a look at the [Community pipelines](custom_pipeline_examples) guide for how to use them and if you're interested in adding a community pipeline check out the [How to contribute a community pipeline](contribute_pipeline) guide!
 
 ## Community components

examples/dreambooth/README.md

@@ -376,18 +376,14 @@ After training, LoRA weights can be loaded very easily into the original pipelin
 load the original pipeline:
 
 ```python
-from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
-import torch
-
-pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
-pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
-pipe.to("cuda")
+from diffusers import DiffusionPipeline
+pipe = DiffusionPipeline.from_pretrained("base-model-name").to("cuda")
 ```
 
-Next, we can load the adapter layers into the UNet with the [`load_attn_procs` function](https://huggingface.co/docs/diffusers/api/loaders#diffusers.loaders.UNet2DConditionLoadersMixin.load_attn_procs).
+Next, we can load the adapter layers into the pipeline with the [`load_lora_weights` function](https://huggingface.co/docs/diffusers/main/en/using-diffusers/loading_adapters#lora).
 
 ```python
-pipe.unet.load_attn_procs("patrickvonplaten/lora_dreambooth_dog_example")
+pipe.load_lora_weights("path-to-the-lora-checkpoint")
 ```
 
 Finally, we can run the model in inference.

examples/t2i_adapter/train_t2i_adapter_sdxl.py

@@ -49,6 +49,7 @@ from diffusers import (
 )
 from diffusers.optimization import get_scheduler
 from diffusers.utils import check_min_version, is_wandb_available
+from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
 from diffusers.utils.import_utils import is_xformers_available
 from diffusers.utils.torch_utils import is_compiled_module
 
@@ -195,7 +196,7 @@ def import_model_class_from_model_name_or_path(
         raise ValueError(f"{model_class} is not supported.")
 
 
-def save_model_card(repo_id: str, image_logs=None, base_model=str, repo_folder=None):
+def save_model_card(repo_id: str, image_logs: dict = None, base_model: str = None, repo_folder: str = None):
     img_str = ""
     if image_logs is not None:
         img_str = "You can find some example images below.\n"
@@ -209,27 +210,25 @@ def save_model_card(repo_id: str, image_logs=None, base_model=str, repo_folder=N
             image_grid(images, 1, len(images)).save(os.path.join(repo_folder, f"images_{i}.png"))
             img_str += f"![images_{i})](./images_{i}.png)\n"
 
-    yaml = f"""
----
-license: creativeml-openrail-m
-base_model: {base_model}
-tags:
-- stable-diffusion-xl
-- stable-diffusion-xl-diffusers
-- text-to-image
-- diffusers
-- t2iadapter
-inference: true
----
-    """
-    model_card = f"""
+    model_description = f"""
 # t2iadapter-{repo_id}
 
 These are t2iadapter weights trained on {base_model} with new type of conditioning.
 {img_str}
 """
-    with open(os.path.join(repo_folder, "README.md"), "w") as f:
-        f.write(yaml + model_card)
+    model_card = load_or_create_model_card(
+        repo_id_or_path=repo_id,
+        from_training=True,
+        license="creativeml-openrail-m",
+        base_model=base_model,
+        model_description=model_description,
+        inference=True,
+    )
+
+    tags = ["stable-diffusion-xl", "stable-diffusion-xl-diffusers", "text-to-image", "diffusers", "t2iadapter"]
+    model_card = populate_model_card(model_card, tags=tags)
+
+    model_card.save(os.path.join(repo_folder, "README.md"))
 
 
 def parse_args(input_args=None):

examples/text_to_image/train_text_to_image.py

@@ -67,8 +67,8 @@ DATASET_NAME_MAPPING = {
 def save_model_card(
     args,
     repo_id: str,
-    images=None,
-    repo_folder=None,
+    images: list = None,
+    repo_folder: str = None,
 ):
     img_str = ""
     if len(images) > 0:

examples/text_to_image/train_text_to_image_lora.py

@@ -56,7 +56,9 @@ check_min_version("0.27.0.dev0")
 logger = get_logger(__name__, log_level="INFO")
 
 
-def save_model_card(repo_id: str, images=None, base_model=str, dataset_name=str, repo_folder=None):
+def save_model_card(
+    repo_id: str, images: list = None, base_model: str = None, dataset_name: str = None, repo_folder: str = None
+):
     img_str = ""
     for i, image in enumerate(images):
         image.save(os.path.join(repo_folder, f"image_{i}.png"))

examples/text_to_image/train_text_to_image_sdxl.py

@@ -66,12 +66,12 @@ DATASET_NAME_MAPPING = {
 
 def save_model_card(
     repo_id: str,
-    images=None,
-    validation_prompt=None,
-    base_model=str,
-    dataset_name=str,
-    repo_folder=None,
-    vae_path=None,
+    images: list = None,
+    validation_prompt: str = None,
+    base_model: str = None,
+    dataset_name: str = None,
+    repo_folder: str = None,
+    vae_path: str = None,
 ):
     img_str = ""
     for i, image in enumerate(images):

scripts/convert_diffusers_to_original_sdxl.py

@@ -167,7 +167,10 @@ vae_conversion_map_attn = [
 
 def reshape_weight_for_sd(w):
     # convert HF linear weights to SD conv2d weights
-    return w.reshape(*w.shape, 1, 1)
+    if not w.ndim == 1:
+        return w.reshape(*w.shape, 1, 1)
+    else:
+        return w
 
 
 def convert_vae_state_dict(vae_state_dict):
@@ -321,11 +324,18 @@ if __name__ == "__main__":
     vae_state_dict = convert_vae_state_dict(vae_state_dict)
     vae_state_dict = {"first_stage_model." + k: v for k, v in vae_state_dict.items()}
 
+    # Convert text encoder 1
     text_enc_dict = convert_openai_text_enc_state_dict(text_enc_dict)
     text_enc_dict = {"conditioner.embedders.0.transformer." + k: v for k, v in text_enc_dict.items()}
 
+    # Convert text encoder 2
     text_enc_2_dict = convert_openclip_text_enc_state_dict(text_enc_2_dict)
     text_enc_2_dict = {"conditioner.embedders.1.model." + k: v for k, v in text_enc_2_dict.items()}
+    # We call the `.T.contiguous()` to match what's done in
+    # https://github.com/huggingface/diffusers/blob/84905ca7287876b925b6bf8e9bb92fec21c78764/src/diffusers/loaders/single_file_utils.py#L1085
+    text_enc_2_dict["conditioner.embedders.1.model.text_projection"] = text_enc_2_dict.pop(
+        "conditioner.embedders.1.model.text_projection.weight"
+    ).T.contiguous()
 
     # Put together new checkpoint
     state_dict = {**unet_state_dict, **vae_state_dict, **text_enc_dict, **text_enc_2_dict}

scripts/convert_diffusers_to_original_stable_diffusion.py

@@ -170,7 +170,10 @@ vae_extra_conversion_map = [
 
 def reshape_weight_for_sd(w):
     # convert HF linear weights to SD conv2d weights
-    return w.reshape(*w.shape, 1, 1)
+    if not w.ndim == 1:
+        return w.reshape(*w.shape, 1, 1)
+    else:
+        return w
 
 
 def convert_vae_state_dict(vae_state_dict):

setup.py

@@ -126,8 +126,8 @@ _deps = [
     "regex!=2019.12.17",
     "requests",
     "tensorboard",
-    "torch>=1.4,<2.2.0",
-    "torchvision<0.17",
+    "torch>=1.4",
+    "torchvision",
     "transformers>=4.25.1",
     "urllib3<=2.0.0",
 ]

src/diffusers/dependency_versions_table.py

@@ -38,8 +38,8 @@ deps = {
     "regex": "regex!=2019.12.17",
     "requests": "requests",
     "tensorboard": "tensorboard",
-    "torch": "torch>=1.4,<2.2.0",
-    "torchvision": "torchvision<0.17",
+    "torch": "torch>=1.4",
+    "torchvision": "torchvision",
     "transformers": "transformers>=4.25.1",
     "urllib3": "urllib3<=2.0.0",
 }

src/diffusers/loaders/single_file_utils.py

@@ -1112,7 +1112,6 @@ def create_text_encoder_from_open_clip_checkpoint(
             text_model_dict[diffusers_key + ".q_proj.bias"] = weight_value[:text_proj_dim]
             text_model_dict[diffusers_key + ".k_proj.bias"] = weight_value[text_proj_dim : text_proj_dim * 2]
             text_model_dict[diffusers_key + ".v_proj.bias"] = weight_value[text_proj_dim * 2 :]
-
         else:
             text_model_dict[diffusers_key] = checkpoint[key]
 

src/diffusers/pipelines/pipeline_utils.py

@@ -981,10 +981,9 @@ class DiffusionPipeline(ConfigMixin, PushToHubMixin):
             revision (`str`, *optional*, defaults to `"main"`):
                 The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
                 allowed by Git.
-            custom_revision (`str`, *optional*, defaults to `"main"`):
+            custom_revision (`str`, *optional*):
                 The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
-                `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
-                custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
+                `revision` when loading a custom pipeline from the Hub. Defaults to the latest stable 🤗 Diffusers version.
             mirror (`str`, *optional*):
                 Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
                 guarantee the timeliness or safety of the source, and you should refer to the mirror site for more

tests/pipelines/pixart_alpha/test_pixart.py

@@ -27,7 +27,13 @@ from diffusers import (
     PixArtAlphaPipeline,
     Transformer2DModel,
 )
-from diffusers.utils.testing_utils import enable_full_determinism, require_torch_gpu, slow, torch_device
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
 
 from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
 from ..test_pipelines_common import PipelineTesterMixin, to_np
@@ -332,37 +338,35 @@ class PixArtAlphaPipelineIntegrationTests(unittest.TestCase):
         torch.cuda.empty_cache()
 
     def test_pixart_1024(self):
-        generator = torch.manual_seed(0)
+        generator = torch.Generator("cpu").manual_seed(0)
 
         pipe = PixArtAlphaPipeline.from_pretrained(self.ckpt_id_1024, torch_dtype=torch.float16)
         pipe.enable_model_cpu_offload()
         prompt = self.prompt
 
-        image = pipe(prompt, generator=generator, output_type="np").images
+        image = pipe(prompt, generator=generator, num_inference_steps=2, output_type="np").images
 
         image_slice = image[0, -3:, -3:, -1]
+        expected_slice = np.array([0.0742, 0.0835, 0.2114, 0.0295, 0.0784, 0.2361, 0.1738, 0.2251, 0.3589])
 
-        expected_slice = np.array([0.1941, 0.2117, 0.2188, 0.1946, 0.218, 0.2124, 0.199, 0.2437, 0.2583])
-
-        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
-        self.assertLessEqual(max_diff, 1e-3)
+        max_diff = numpy_cosine_similarity_distance(image_slice.flatten(), expected_slice)
+        self.assertLessEqual(max_diff, 1e-4)
 
     def test_pixart_512(self):
-        generator = torch.manual_seed(0)
+        generator = torch.Generator("cpu").manual_seed(0)
 
         pipe = PixArtAlphaPipeline.from_pretrained(self.ckpt_id_512, torch_dtype=torch.float16)
         pipe.enable_model_cpu_offload()
 
         prompt = self.prompt
 
-        image = pipe(prompt, generator=generator, output_type="np").images
+        image = pipe(prompt, generator=generator, num_inference_steps=2, output_type="np").images
 
         image_slice = image[0, -3:, -3:, -1]
+        expected_slice = np.array([0.3477, 0.3882, 0.4541, 0.3413, 0.3821, 0.4463, 0.4001, 0.4409, 0.4958])
 
-        expected_slice = np.array([0.2637, 0.291, 0.2939, 0.207, 0.2512, 0.2783, 0.2168, 0.2324, 0.2817])
-
-        max_diff = np.abs(image_slice.flatten() - expected_slice).max()
-        self.assertLessEqual(max_diff, 1e-3)
+        max_diff = numpy_cosine_similarity_distance(image_slice.flatten(), expected_slice)
+        self.assertLessEqual(max_diff, 1e-4)
 
     def test_pixart_1024_without_resolution_binning(self):
         generator = torch.manual_seed(0)
@@ -372,7 +376,7 @@ class PixArtAlphaPipelineIntegrationTests(unittest.TestCase):
 
         prompt = self.prompt
         height, width = 1024, 768
-        num_inference_steps = 10
+        num_inference_steps = 2
 
         image = pipe(
             prompt,
@@ -406,7 +410,7 @@ class PixArtAlphaPipelineIntegrationTests(unittest.TestCase):
 
         prompt = self.prompt
         height, width = 512, 768
-        num_inference_steps = 10
+        num_inference_steps = 2
 
         image = pipe(
             prompt,