update

Author: a-r-r-o-w

scripts/convert_hunyuan_video_to_diffusers.py

@@ -3,11 +3,19 @@
 
 import torch
 from accelerate import init_empty_weights
-from transformers import AutoModel, AutoTokenizer, CLIPTextModel, CLIPTokenizer, LlavaForConditionalGeneration
+from transformers import (
+    AutoModel,
+    AutoTokenizer,
+    CLIPImageProcessor,
+    CLIPTextModel,
+    CLIPTokenizer,
+    LlavaForConditionalGeneration,
+)
 
 from diffusers import (
     AutoencoderKLHunyuanVideo,
     FlowMatchEulerDiscreteScheduler,
+    HunyuanVideoImageToVideoPipeline,
     HunyuanVideoPipeline,
     HunyuanVideoTransformer3DModel,
 )
@@ -153,7 +161,7 @@ def remap_single_transformer_blocks_(key, state_dict):
         "rope_theta": 256.0,
         "rope_axes_dim": (16, 56, 56),
     },
-    "HYVideo-T/2": {
+    "HYVideo-T/2-I2V": {
         "in_channels": 16 * 2 + 1,
         "out_channels": 16,
         "num_attention_heads": 24,
@@ -286,23 +294,39 @@ def get_args():
     if args.save_pipeline:
         if args.transformer_type == "HYVideo-T/2-cfgdistill":
             text_encoder = AutoModel.from_pretrained(args.text_encoder_path, torch_dtype=torch.float16)
+            tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_path, padding_side="right")
+            text_encoder_2 = CLIPTextModel.from_pretrained(args.text_encoder_2_path, torch_dtype=torch.float16)
+            tokenizer_2 = CLIPTokenizer.from_pretrained(args.text_encoder_2_path)
+            scheduler = FlowMatchEulerDiscreteScheduler(shift=args.flow_shift)
+
+            pipe = HunyuanVideoPipeline(
+                transformer=transformer,
+                vae=vae,
+                text_encoder=text_encoder,
+                tokenizer=tokenizer,
+                text_encoder_2=text_encoder_2,
+                tokenizer_2=tokenizer_2,
+                scheduler=scheduler,
+            )
+            pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
         else:
             text_encoder = LlavaForConditionalGeneration.from_pretrained(
                 args.text_encoder_path, torch_dtype=torch.float16
             )
-
-        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_path, padding_side="right")
-        text_encoder_2 = CLIPTextModel.from_pretrained(args.text_encoder_2_path, torch_dtype=torch.float16)
-        tokenizer_2 = CLIPTokenizer.from_pretrained(args.text_encoder_2_path)
-        scheduler = FlowMatchEulerDiscreteScheduler(shift=args.flow_shift)
-
-        pipe = HunyuanVideoPipeline(
-            transformer=transformer,
-            vae=vae,
-            text_encoder=text_encoder,
-            tokenizer=tokenizer,
-            text_encoder_2=text_encoder_2,
-            tokenizer_2=tokenizer_2,
-            scheduler=scheduler,
-        )
-        pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")
+            tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_path, padding_side="right")
+            text_encoder_2 = CLIPTextModel.from_pretrained(args.text_encoder_2_path, torch_dtype=torch.float16)
+            tokenizer_2 = CLIPTokenizer.from_pretrained(args.text_encoder_2_path)
+            scheduler = FlowMatchEulerDiscreteScheduler(shift=args.flow_shift)
+            image_processor = CLIPImageProcessor.from_pretrained(args.text_encoder_2_path)
+
+            pipe = HunyuanVideoImageToVideoPipeline(
+                transformer=transformer,
+                vae=vae,
+                text_encoder=text_encoder,
+                tokenizer=tokenizer,
+                text_encoder_2=text_encoder_2,
+                tokenizer_2=tokenizer_2,
+                scheduler=scheduler,
+                image_processor=image_processor,
+            )
+            pipe.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB")

src/diffusers/__init__.py

@@ -313,6 +313,7 @@
             "HunyuanDiTPAGPipeline",
             "HunyuanDiTPipeline",
             "HunyuanSkyreelsImageToVideoPipeline",
+            "HunyuanVideoImageToVideoPipeline",
             "HunyuanVideoPipeline",
             "I2VGenXLPipeline",
             "IFImg2ImgPipeline",
@@ -823,6 +824,7 @@
             HunyuanDiTPAGPipeline,
             HunyuanDiTPipeline,
             HunyuanSkyreelsImageToVideoPipeline,
+            HunyuanVideoImageToVideoPipeline,
             HunyuanVideoPipeline,
             I2VGenXLPipeline,
             IFImg2ImgPipeline,

src/diffusers/pipelines/__init__.py

@@ -222,7 +222,11 @@
         "EasyAnimateControlPipeline",
     ]
     _import_structure["hunyuandit"] = ["HunyuanDiTPipeline"]
-    _import_structure["hunyuan_video"] = ["HunyuanVideoPipeline", "HunyuanSkyreelsImageToVideoPipeline"]
+    _import_structure["hunyuan_video"] = [
+        "HunyuanVideoPipeline",
+        "HunyuanSkyreelsImageToVideoPipeline",
+        "HunyuanVideoImageToVideoPipeline",
+    ]
     _import_structure["kandinsky"] = [
         "KandinskyCombinedPipeline",
         "KandinskyImg2ImgCombinedPipeline",
@@ -570,7 +574,11 @@
             FluxPriorReduxPipeline,
             ReduxImageEncoder,
         )
-        from .hunyuan_video import HunyuanSkyreelsImageToVideoPipeline, HunyuanVideoPipeline
+        from .hunyuan_video import (
+            HunyuanSkyreelsImageToVideoPipeline,
+            HunyuanVideoImageToVideoPipeline,
+            HunyuanVideoPipeline,
+        )
         from .hunyuandit import HunyuanDiTPipeline
         from .i2vgen_xl import I2VGenXLPipeline
         from .kandinsky import (

src/diffusers/pipelines/hunyuan_video/__init__.py

@@ -24,6 +24,7 @@
 else:
     _import_structure["pipeline_hunyuan_skyreels_image2video"] = ["HunyuanSkyreelsImageToVideoPipeline"]
     _import_structure["pipeline_hunyuan_video"] = ["HunyuanVideoPipeline"]
+    _import_structure["pipeline_hunyuan_video_image2video"] = ["HunyuanVideoImageToVideoPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
@@ -35,6 +36,7 @@
     else:
         from .pipeline_hunyuan_skyreels_image2video import HunyuanSkyreelsImageToVideoPipeline
         from .pipeline_hunyuan_video import HunyuanVideoPipeline
+        from .pipeline_hunyuan_video_image2video import HunyuanVideoImageToVideoPipeline
 
 else:
     import sys