Merge branch 'main' into enable-hotswap-testing-ci

sayakpaul · web-flow · commit 580e7aebf8df · 2025-04-21T14:36:39.000+05:30
diff --git a/docs/source/en/api/pipelines/wan.md b/docs/source/en/api/pipelines/wan.md
@@ -133,6 +133,60 @@ output = pipe(
 export_to_video(output, "wan-i2v.mp4", fps=16)
 ```
 
+### First and Last Frame Interpolation
+
+```python
+import numpy as np
+import torch
+import torchvision.transforms.functional as TF
+from diffusers import AutoencoderKLWan, WanImageToVideoPipeline
+from diffusers.utils import export_to_video, load_image
+from transformers import CLIPVisionModel
+
+
+model_id = "Wan-AI/Wan2.1-FLF2V-14B-720P-diffusers"
+image_encoder = CLIPVisionModel.from_pretrained(model_id, subfolder="image_encoder", torch_dtype=torch.float32)
+vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
+pipe = WanImageToVideoPipeline.from_pretrained(
+    model_id, vae=vae, image_encoder=image_encoder, torch_dtype=torch.bfloat16
+)
+pipe.to("cuda")
+
+first_frame = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png")
+last_frame = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png")
+
+def aspect_ratio_resize(image, pipe, max_area=720 * 1280):
+    aspect_ratio = image.height / image.width
+    mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
+    height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
+    width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
+    image = image.resize((width, height))
+    return image, height, width
+
+def center_crop_resize(image, height, width):
+    # Calculate resize ratio to match first frame dimensions
+    resize_ratio = max(width / image.width, height / image.height)
+    
+    # Resize the image
+    width = round(image.width * resize_ratio)
+    height = round(image.height * resize_ratio)
+    size = [width, height]
+    image = TF.center_crop(image, size)
+    
+    return image, height, width
+
+first_frame, height, width = aspect_ratio_resize(first_frame, pipe)
+if last_frame.size != first_frame.size:
+    last_frame, _, _ = center_crop_resize(last_frame, height, width)
+
+prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective."
+
+output = pipe(
+    image=first_frame, last_image=last_frame, prompt=prompt, height=height, width=width, guidance_scale=5.5
+).frames[0]
+export_to_video(output, "output.mp4", fps=16)
+```
+
 ### Video to Video Generation
 
 ```python
diff --git a/examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py b/examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py
@@ -1915,17 +1915,22 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
             free_memory()
 
     # 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)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * accelerator.num_processes * num_update_steps_per_epoch
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )
@@ -1949,7 +1954,6 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                 lr_scheduler,
             )
         else:
-            print("I SHOULD BE HERE")
             transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
                 transformer, text_encoder_one, optimizer, train_dataloader, lr_scheduler
             )
@@ -1961,8 +1965,14 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 
diff --git a/examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py b/examples/community/pipeline_stable_diffusion_xl_controlnet_adapter_inpaint.py
@@ -33,7 +33,6 @@
 from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
 from diffusers.loaders import (
     FromSingleFileMixin,
-    StableDiffusionLoraLoaderMixin,
     StableDiffusionXLLoraLoaderMixin,
     TextualInversionLoaderMixin,
 )
@@ -300,7 +299,7 @@ def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
 
 
 class StableDiffusionXLControlNetAdapterInpaintPipeline(
-    DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin, StableDiffusionLoraLoaderMixin
+    DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin
 ):
     r"""
     Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter
diff --git a/examples/dreambooth/train_dreambooth_flux.py b/examples/dreambooth/train_dreambooth_flux.py
@@ -1407,17 +1407,22 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                 tokens_two = torch.cat([tokens_two, class_tokens_two], dim=0)
 
     # 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)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * accelerator.num_processes * num_update_steps_per_epoch
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )
@@ -1444,8 +1449,14 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 
diff --git a/examples/dreambooth/train_dreambooth_lora_flux.py b/examples/dreambooth/train_dreambooth_lora_flux.py
@@ -1524,17 +1524,22 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
             free_memory()
 
     # 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)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * accelerator.num_processes * num_update_steps_per_epoch
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )
@@ -1561,8 +1566,14 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 
diff --git a/examples/dreambooth/train_dreambooth_lora_sdxl.py b/examples/dreambooth/train_dreambooth_lora_sdxl.py
@@ -1523,17 +1523,22 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
                 tokens_two = torch.cat([tokens_two, class_tokens_two], dim=0)
 
     # 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)
+    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
+    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
     if args.max_train_steps is None:
-        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
-        overrode_max_train_steps = True
+        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
+        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
+        num_training_steps_for_scheduler = (
+            args.num_train_epochs * accelerator.num_processes * num_update_steps_per_epoch
+        )
+    else:
+        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes
 
     lr_scheduler = get_scheduler(
         args.lr_scheduler,
         optimizer=optimizer,
-        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
-        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_warmup_steps=num_warmup_steps_for_scheduler,
+        num_training_steps=num_training_steps_for_scheduler,
         num_cycles=args.lr_num_cycles,
         power=args.lr_power,
     )
@@ -1550,7 +1555,14 @@ def compute_text_embeddings(prompt, text_encoders, tokenizers):
 
     # We need to recalculate our total training steps as the size of the training dataloader may have changed.
     num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
-    if overrode_max_train_steps:
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        if num_training_steps_for_scheduler != args.max_train_steps:
+            logger.warning(
+                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
+                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
+                f"This inconsistency may result in the learning rate scheduler not functioning properly."
+            )
         args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
     # Afterwards we recalculate our number of training epochs
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
diff --git a/scripts/convert_wan_to_diffusers.py b/scripts/convert_wan_to_diffusers.py
@@ -39,6 +39,24 @@
     "img_emb.proj.1": "condition_embedder.image_embedder.ff.net.0.proj",
     "img_emb.proj.3": "condition_embedder.image_embedder.ff.net.2",
     "img_emb.proj.4": "condition_embedder.image_embedder.norm2",
+    # for the FLF2V model
+    "img_emb.emb_pos": "condition_embedder.image_embedder.pos_embed",
+    # Add attention component mappings
+    "self_attn.q": "attn1.to_q",
+    "self_attn.k": "attn1.to_k",
+    "self_attn.v": "attn1.to_v",
+    "self_attn.o": "attn1.to_out.0",
+    "self_attn.norm_q": "attn1.norm_q",
+    "self_attn.norm_k": "attn1.norm_k",
+    "cross_attn.q": "attn2.to_q",
+    "cross_attn.k": "attn2.to_k",
+    "cross_attn.v": "attn2.to_v",
+    "cross_attn.o": "attn2.to_out.0",
+    "cross_attn.norm_q": "attn2.norm_q",
+    "cross_attn.norm_k": "attn2.norm_k",
+    "attn2.to_k_img": "attn2.add_k_proj",
+    "attn2.to_v_img": "attn2.add_v_proj",
+    "attn2.norm_k_img": "attn2.norm_added_k",
 }
 
 TRANSFORMER_SPECIAL_KEYS_REMAP = {}
@@ -135,6 +153,28 @@ def get_transformer_config(model_type: str) -> Dict[str, Any]:
                 "text_dim": 4096,
             },
         }
+    elif model_type == "Wan-FLF2V-14B-720P":
+        config = {
+            "model_id": "ypyp/Wan2.1-FLF2V-14B-720P",  # This is just a placeholder
+            "diffusers_config": {
+                "image_dim": 1280,
+                "added_kv_proj_dim": 5120,
+                "attention_head_dim": 128,
+                "cross_attn_norm": True,
+                "eps": 1e-06,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "in_channels": 36,
+                "num_attention_heads": 40,
+                "num_layers": 40,
+                "out_channels": 16,
+                "patch_size": [1, 2, 2],
+                "qk_norm": "rms_norm_across_heads",
+                "text_dim": 4096,
+                "rope_max_seq_len": 1024,
+                "pos_embed_seq_len": 257 * 2,
+            },
+        }
     return config
 
 
@@ -393,11 +433,12 @@ def get_args():
     vae = convert_vae()
     text_encoder = UMT5EncoderModel.from_pretrained("google/umt5-xxl")
     tokenizer = AutoTokenizer.from_pretrained("google/umt5-xxl")
+    flow_shift = 16.0 if "FLF2V" in args.model_type else 3.0
     scheduler = UniPCMultistepScheduler(
-        prediction_type="flow_prediction", use_flow_sigmas=True, num_train_timesteps=1000, flow_shift=3.0
+        prediction_type="flow_prediction", use_flow_sigmas=True, num_train_timesteps=1000, flow_shift=flow_shift
     )
 
-    if "I2V" in args.model_type:
+    if "I2V" in args.model_type or "FLF2V" in args.model_type:
         image_encoder = CLIPVisionModelWithProjection.from_pretrained(
             "laion/CLIP-ViT-H-14-laion2B-s32B-b79K", torch_dtype=torch.bfloat16
         )
diff --git a/src/diffusers/models/transformers/transformer_hidream_image.py b/src/diffusers/models/transformers/transformer_hidream_image.py
@@ -918,5 +918,5 @@ def forward(
             unscale_lora_layers(self, lora_scale)
 
         if not return_dict:
-            return (output, hidden_states_masks)
-        return Transformer2DModelOutput(sample=output, mask=hidden_states_masks)
+            return (output,)
+        return Transformer2DModelOutput(sample=output)
diff --git a/src/diffusers/models/transformers/transformer_wan.py b/src/diffusers/models/transformers/transformer_wan.py
diff --git a/src/diffusers/pipelines/wan/pipeline_wan_i2v.py b/src/diffusers/pipelines/wan/pipeline_wan_i2v.py
diff --git a/tests/pipelines/wan/test_wan_image_to_video.py b/tests/pipelines/wan/test_wan_image_to_video.py
