update

Author: DN6

src/diffusers/loaders/single_file_model.py

@@ -22,6 +22,7 @@
 from ..utils import deprecate, is_accelerate_available, logging
 from .single_file_utils import (
     SingleFileComponentError,
+    convert_animatediff_checkpoint_to_diffusers,
     convert_controlnet_checkpoint,
     convert_ldm_unet_checkpoint,
     convert_ldm_vae_checkpoint,
@@ -70,6 +71,9 @@
         "checkpoint_mapping_fn": convert_sd3_transformer_checkpoint_to_diffusers,
         "default_subfolder": "transformer",
     },
+    "MotionAdapter": {
+        "checkpoint_mapping_fn": convert_animatediff_checkpoint_to_diffusers,
+    },
 }
 
 

src/diffusers/loaders/single_file_utils.py

@@ -74,6 +74,9 @@
     "stable_cascade_stage_b": "down_blocks.1.0.channelwise.0.weight",
     "stable_cascade_stage_c": "clip_txt_mapper.weight",
     "sd3": "model.diffusion_model.joint_blocks.0.context_block.adaLN_modulation.1.bias",
+    "animatediff": "down_blocks.0.motion_modules.0.temporal_transformer.norm.weight",
+    "animatediff_v2": "mid_block.motion_modules.0.temporal_transformer.norm.bias",
+    "animatediff_sdxl_beta": "down_blocks.3.motion_modules.0.temporal_transformer.norm.bias",
 }
 
 DIFFUSERS_DEFAULT_PIPELINE_PATHS = {
@@ -103,6 +106,9 @@
     "sd3": {
         "pretrained_model_name_or_path": "stabilityai/stable-diffusion-3-medium-diffusers",
     },
+    "animatediff_v2": "guoyww/animatediff-motion-adapter-v1-5-2",
+    "animatediff_v3": "guoyww/animatediff-motion-adapter-v1-5-3",
+    "animatediff_sdxl": "guoyww/animatediff-motion-adapter-sdxl-beta",
 }
 
 # Use to configure model sample size when original config is provided
@@ -485,6 +491,13 @@ def infer_diffusers_model_type(checkpoint):
     elif CHECKPOINT_KEY_NAMES["sd3"] in checkpoint:
         model_type = "sd3"
 
+    elif CHECKPOINT_KEY_NAMES["animatediff"] in checkpoint:
+        if CHECKPOINT_KEY_NAMES["animatediff_v2"] in checkpoint:
+            model_type = "animatediff_v2"
+        elif CHECKPOINT_KEY_NAMES["animatediff_sdxl_beta"] in checkpoint:
+            model_type = "animatediff_sdxl_beta"
+        else:
+            model_type = "animatediff_v3"
     else:
         model_type = "v1"
 
@@ -1822,3 +1835,22 @@ def create_diffusers_t5_model_from_checkpoint(
                 param.data = param.data.to(torch.float32)
 
     return model
+
+
+def convert_animatediff_checkpoint_to_diffusers(checkpoint, **kwargs):
+    converted_state_dict = {}
+    for k, v in checkpoint.items():
+        if "pos_encoder" in k:
+            continue
+
+        else:
+            converted_state_dict[
+                k.replace(".norms.0", ".norm1")
+                .replace(".norms.1", ".norm2")
+                .replace(".ff_norm", ".norm3")
+                .replace(".attention_blocks.0", ".attn1")
+                .replace(".attention_blocks.1", ".attn2")
+                .replace(".temporal_transformer", "")
+            ] = v
+
+    return converted_state_dict

src/diffusers/models/unets/unet_motion_model.py

@@ -19,7 +19,7 @@
 import torch.utils.checkpoint
 
 from ...configuration_utils import ConfigMixin, FrozenDict, register_to_config
-from ...loaders import UNet2DConditionLoadersMixin
+from ...loaders import FromOriginalModelMixin, UNet2DConditionLoadersMixin
 from ...utils import logging
 from ..attention_processor import (
     ADDED_KV_ATTENTION_PROCESSORS,
@@ -93,7 +93,7 @@ def __init__(
             )
 
 
-class MotionAdapter(ModelMixin, ConfigMixin):
+class MotionAdapter(ModelMixin, ConfigMixin, FromOriginalModelMixin):
     @register_to_config
     def __init__(
         self,