initial

docs/source/en/api/cache.md

@@ -29,7 +29,7 @@ Cache methods speedup diffusion transformers by storing and reusing intermediate
 
 [[autodoc]] apply_faster_cache
 
-## FirstBlockCacheConfig
+### FirstBlockCacheConfig
 
 [[autodoc]] FirstBlockCacheConfig
 

docs/source/en/optimization/cache.md

@@ -68,20 +68,6 @@ config = FasterCacheConfig(
 pipeline.transformer.enable_cache(config)
 ```
 
-## FirstBlockCache
-
-[FirstBlock Cache](https://huggingface.co/docs/diffusers/main/en/api/cache#diffusers.FirstBlockCacheConfig) checks how much the early layers of the denoiser changes from one timestep to the next. If the change is small, the model skips the expensive later layers and reuses the previous output.
-
-```py
-import torch
-from diffusers import DiffusionPipeline
-from diffusers.hooks import apply_first_block_cache, FirstBlockCacheConfig
-
-pipeline = DiffusionPipeline.from_pretrained(
-    "Qwen/Qwen-Image", torch_dtype=torch.bfloat16
-)
-apply_first_block_cache(pipeline.transformer, FirstBlockCacheConfig(threshold=0.2))
-```
 ## TaylorSeer Cache
 
 [TaylorSeer Cache](https://huggingface.co/papers/2403.06923) accelerates diffusion inference by using Taylor series expansions to approximate and cache intermediate activations across denoising steps. The method predicts future outputs based on past computations, reusing them at specified intervals to reduce redundant calculations.
@@ -101,7 +87,8 @@ from diffusers import FluxPipeline, TaylorSeerCacheConfig
 pipe = FluxPipeline.from_pretrained(
     "black-forest-labs/FLUX.1-dev",
     torch_dtype=torch.bfloat16,
-).to("cuda")
+)
+pipe.to("cuda")
 
 config = TaylorSeerCacheConfig(
     cache_interval=5,
@@ -110,4 +97,4 @@ config = TaylorSeerCacheConfig(
     taylor_factors_dtype=torch.bfloat16,
 )
 pipe.transformer.enable_cache(config)
-```
+```

src/diffusers/models/cache_utils.py

@@ -41,11 +41,9 @@ class CacheMixin:
         Enable caching techniques on the model.
 
         Args:
-            config (`Union[PyramidAttentionBroadcastConfig, FasterCacheConfig, FirstBlockCacheConfig]`):
+            config (`Union[PyramidAttentionBroadcastConfig]`):
                 The configuration for applying the caching technique. Currently supported caching techniques are:
                     - [`~hooks.PyramidAttentionBroadcastConfig`]
-                    - [`~hooks.FasterCacheConfig`]
-                    - [`~hooks.FirstBlockCacheConfig`]
 
         Example:
 

src/diffusers/modular_pipelines/mellon_node_utils.py

@@ -68,6 +68,10 @@ class MellonParam:
     def image_latents(cls, display: str = "input") -> "MellonParam":
         return cls(name="image_latents", label="Image Latents", type="latents", display=display)
 
+    @classmethod
+    def first_frame_latents(cls, display: str = "input") -> "MellonParam":
+        return cls(name="first_frame_latents", label="First Frame Latents", type="latents", display=display)
+
     @classmethod
     def image_latents_with_strength(cls) -> "MellonParam":
         return cls(
@@ -89,6 +93,10 @@ class MellonParam:
     def embeddings(cls, display: str = "output") -> "MellonParam":
         return cls(name="embeddings", label="Text Embeddings", type="embeddings", display=display)
 
+    @classmethod
+    def image_embeds(cls, display: str = "output") -> "MellonParam":
+        return cls(name="image_embeds", label="Image Embeddings", type="image_embeds", display=display)
+
     @classmethod
     def controlnet_conditioning_scale(cls, default: float = 0.5) -> "MellonParam":
         return cls(
@@ -186,6 +194,16 @@ class MellonParam:
         """
         return cls(name="vae", label="VAE", type="diffusers_auto_model", display="input")
 
+    @classmethod
+    def image_encoder(cls) -> "MellonParam":
+        """
+        Image Encoder model info dict.
+
+        Contains keys like 'model_id', 'repo_id', 'execution_device' etc. Use components.get_one(model_id) to retrieve
+        the actual model.
+        """
+        return cls(name="image_encoder", label="Image Encoder", type="diffusers_auto_model", display="input")
+
     @classmethod
     def unet(cls) -> "MellonParam":
         """

src/diffusers/modular_pipelines/wan/modular_blocks.py

@@ -84,7 +84,7 @@ class WanImage2VideoImageEncoderStep(SequentialPipelineBlocks):
 class WanImage2VideoVaeImageEncoderStep(SequentialPipelineBlocks):
     model_name = "wan"
     block_classes = [WanImageResizeStep, WanVaeImageEncoderStep]
-    block_names = ["image_resize", "vae_image_encoder"]
+    block_names = ["image_resize", "vae_encoder"]
 
     @property
     def description(self):
@@ -142,7 +142,7 @@ class WanFLF2VImageEncoderStep(SequentialPipelineBlocks):
 class WanFLF2VVaeImageEncoderStep(SequentialPipelineBlocks):
     model_name = "wan"
     block_classes = [WanImageResizeStep, WanImageCropResizeStep, WanFirstLastFrameVaeImageEncoderStep]
-    block_names = ["image_resize", "last_image_resize", "vae_image_encoder"]
+    block_names = ["image_resize", "last_image_resize", "vae_encoder"]
 
     @property
     def description(self):
@@ -203,7 +203,7 @@ class WanAutoImageEncoderStep(AutoPipelineBlocks):
 ## vae encoder
 class WanAutoVaeImageEncoderStep(AutoPipelineBlocks):
     block_classes = [WanFLF2VVaeImageEncoderStep, WanImage2VideoVaeImageEncoderStep]
-    block_names = ["flf2v_vae_image_encoder", "image2video_vae_image_encoder"]
+    block_names = ["flf2v_vae_encoder", "image2video_vae_encoder"]
     block_trigger_inputs = ["last_image", "image"]
 
     @property
@@ -251,7 +251,7 @@ class WanAutoBlocks(SequentialPipelineBlocks):
     block_names = [
         "text_encoder",
         "image_encoder",
-        "vae_image_encoder",
+        "vae_encoder",
         "denoise",
         "decode",
     ]
@@ -353,7 +353,7 @@ class Wan22AutoBlocks(SequentialPipelineBlocks):
     ]
     block_names = [
         "text_encoder",
-        "vae_image_encoder",
+        "vae_encoder",
         "denoise",
         "decode",
     ]
@@ -384,7 +384,7 @@ IMAGE2VIDEO_BLOCKS = InsertableDict(
     [
         ("image_resize", WanImageResizeStep),
         ("image_encoder", WanImage2VideoImageEncoderStep),
-        ("vae_image_encoder", WanImage2VideoVaeImageEncoderStep),
+        ("vae_encoder", WanImage2VideoVaeImageEncoderStep),
         ("input", WanTextInputStep),
         ("additional_inputs", WanAdditionalInputsStep(image_latent_inputs=["first_frame_latents"])),
         ("set_timesteps", WanSetTimestepsStep),
@@ -401,7 +401,7 @@ FLF2V_BLOCKS = InsertableDict(
         ("image_resize", WanImageResizeStep),
         ("last_image_resize", WanImageCropResizeStep),
         ("image_encoder", WanFLF2VImageEncoderStep),
-        ("vae_image_encoder", WanFLF2VVaeImageEncoderStep),
+        ("vae_encoder", WanFLF2VVaeImageEncoderStep),
         ("input", WanTextInputStep),
         ("additional_inputs", WanAdditionalInputsStep(image_latent_inputs=["first_last_frame_latents"])),
         ("set_timesteps", WanSetTimestepsStep),
@@ -416,7 +416,7 @@ AUTO_BLOCKS = InsertableDict(
     [
         ("text_encoder", WanTextEncoderStep),
         ("image_encoder", WanAutoImageEncoderStep),
-        ("vae_image_encoder", WanAutoVaeImageEncoderStep),
+        ("vae_encoder", WanAutoVaeImageEncoderStep),
         ("denoise", WanAutoDenoiseStep),
         ("decode", WanImageVaeDecoderStep),
     ]
@@ -438,7 +438,7 @@ TEXT2VIDEO_BLOCKS_WAN22 = InsertableDict(
 IMAGE2VIDEO_BLOCKS_WAN22 = InsertableDict(
     [
         ("image_resize", WanImageResizeStep),
-        ("vae_image_encoder", WanImage2VideoVaeImageEncoderStep),
+        ("vae_encoder", WanImage2VideoVaeImageEncoderStep),
         ("input", WanTextInputStep),
         ("set_timesteps", WanSetTimestepsStep),
         ("prepare_latents", WanPrepareLatentsStep),
@@ -450,7 +450,7 @@ IMAGE2VIDEO_BLOCKS_WAN22 = InsertableDict(
 AUTO_BLOCKS_WAN22 = InsertableDict(
     [
         ("text_encoder", WanTextEncoderStep),
-        ("vae_image_encoder", WanAutoVaeImageEncoderStep),
+        ("vae_encoder", WanAutoVaeImageEncoderStep),
         ("denoise", Wan22AutoDenoiseStep),
         ("decode", WanImageVaeDecoderStep),
     ]