update

* fix: bias loading bug

* fixes for SDXL

* apply changes to the conversion script to match single_file_utils.py

* do transpose to match the single file loading logic.

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):

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):

src/diffusers/loaders/autoencoder.py

@@ -38,6 +38,9 @@ class FromOriginalVAEMixin:
                     - A link to the `.ckpt` file (for example
                       `"https://huggingface.co/<repo_id>/blob/main/<path_to_file>.ckpt"`) on the Hub.
                     - A path to a *file* containing all pipeline weights.
+            config_file (`str`, *optional*):
+                Filepath to the configuration YAML file associated with the model. If not provided it will default to:
+                https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml
             torch_dtype (`str` or `torch.dtype`, *optional*):
                 Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
                 dtype is automatically derived from the model's weights.
@@ -65,6 +68,13 @@ class FromOriginalVAEMixin:
             image_size (`int`, *optional*, defaults to 512):
                 The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
                 Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+            scaling_factor (`float`, *optional*, defaults to 0.18215):
+                The component-wise standard deviation of the trained latent space computed using the first batch of the
+                training set. This is used to scale the latent space to have unit variance when training the diffusion
+                model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+                diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z
+                = 1 / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution
+                Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
             use_safetensors (`bool`, *optional*, defaults to `None`):
                 If set to `None`, the safetensors weights are downloaded if they're available **and** if the
                 safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
@@ -92,6 +102,7 @@ class FromOriginalVAEMixin:
         """
 
         original_config_file = kwargs.pop("original_config_file", None)
+        config_file = kwargs.pop("config_file", None)
         resume_download = kwargs.pop("resume_download", False)
         force_download = kwargs.pop("force_download", False)
         proxies = kwargs.pop("proxies", None)
@@ -103,6 +114,13 @@ class FromOriginalVAEMixin:
         use_safetensors = kwargs.pop("use_safetensors", True)
 
         class_name = cls.__name__
+
+        if (config_file is not None) and (original_config_file is not None):
+            raise ValueError(
+                "You cannot pass both `config_file` and `original_config_file` to `from_single_file`. Please use only one of these arguments."
+            )
+
+        original_config_file = original_config_file or config_file
         original_config, checkpoint = fetch_ldm_config_and_checkpoint(
             pretrained_model_link_or_path=pretrained_model_link_or_path,
             class_name=class_name,
@@ -118,7 +136,10 @@ class FromOriginalVAEMixin:
         )
 
         image_size = kwargs.pop("image_size", None)
-        component = create_diffusers_vae_model_from_ldm(class_name, original_config, checkpoint, image_size=image_size)
+        scaling_factor = kwargs.pop("scaling_factor", None)
+        component = create_diffusers_vae_model_from_ldm(
+            class_name, original_config, checkpoint, image_size=image_size, scaling_factor=scaling_factor
+        )
         vae = component["vae"]
         if torch_dtype is not None:
             vae = vae.to(torch_dtype)

src/diffusers/loaders/single_file_utils.py

@@ -175,6 +175,7 @@ DIFFUSERS_TO_LDM_MAPPING = {
 }
 
 LDM_VAE_KEY = "first_stage_model."
+LDM_VAE_DEFAULT_SCALING_FACTOR = 0.18215
 LDM_UNET_KEY = "model.diffusion_model."
 LDM_CONTROLNET_KEY = "control_model."
 LDM_CLIP_PREFIX_TO_REMOVE = ["cond_stage_model.transformer.", "conditioner.embedders.0.transformer."]
@@ -518,7 +519,10 @@ def create_vae_diffusers_config(original_config, image_size, scaling_factor=None
     Creates a config for the diffusers based on the config of the LDM model.
     """
     vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"]
-    scaling_factor = scaling_factor or original_config["model"]["params"]["scale_factor"]
+    if scaling_factor is None and "scale_factor" in original_config["model"]["params"]:
+        scaling_factor = original_config["model"]["params"]["scale_factor"]
+    elif scaling_factor is None:
+        scaling_factor = LDM_VAE_DEFAULT_SCALING_FACTOR
 
     block_out_channels = [vae_params["ch"] * mult for mult in vae_params["ch_mult"]]
     down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels)
@@ -1112,7 +1116,6 @@ 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]
 
@@ -1174,7 +1177,7 @@ def create_diffusers_unet_model_from_ldm(
 
 
 def create_diffusers_vae_model_from_ldm(
-    pipeline_class_name, original_config, checkpoint, image_size=None, scaling_factor=0.18125
+    pipeline_class_name, original_config, checkpoint, image_size=None, scaling_factor=None
 ):
     # import here to avoid circular imports
     from ..models import AutoencoderKL