Merge pull request #436 from mi804/hunyuanvideo_i2v

support hunyuanvideo-i2v

README.md

@@ -19,7 +19,7 @@ Until now, DiffSynth Studio has supported the following models:
 
 * [Wan-Video](https://github.com/Wan-Video/Wan2.1)
 * [StepVideo](https://github.com/stepfun-ai/Step-Video-T2V)
-* [HunyuanVideo](https://github.com/Tencent/HunyuanVideo)
+* [HunyuanVideo](https://github.com/Tencent/HunyuanVideo), [HunyuanVideo-I2V]()
 * [CogVideoX](https://huggingface.co/THUDM/CogVideoX-5b)
 * [FLUX](https://huggingface.co/black-forest-labs/FLUX.1-dev)
 * [ExVideo](https://huggingface.co/ECNU-CILab/ExVideo-SVD-128f-v1)
@@ -36,6 +36,7 @@ Until now, DiffSynth Studio has supported the following models:
 * [Stable Diffusion](https://huggingface.co/runwayml/stable-diffusion-v1-5)
 
 ## News
+- **March 25, 2025** We support HunyuanVideo-I2V, the image-to-video generation version of HunyuanVideo open-sourced by Tencent. Please refer to [./examples/HunyuanVideo/](./examples/HunyuanVideo/) for more details.
 
 - **February 25, 2025** We support Wan-Video, a collection of SOTA video synthesis models open-sourced by Alibaba. See [./examples/wanvideo/](./examples/wanvideo/).
 

diffsynth/configs/model_config.py

@@ -135,6 +135,7 @@ huggingface_model_loader_configs = [
     ("CogVideoXTransformer3DModel", "diffsynth.models.cog_dit", "cog_dit", "CogDiT"),
     ("SiglipModel", "transformers.models.siglip.modeling_siglip", "siglip_vision_model", "SiglipVisionModel"),
     ("LlamaForCausalLM", "diffsynth.models.hunyuan_video_text_encoder", "hunyuan_video_text_encoder_2", "HunyuanVideoLLMEncoder"),
+    ("LlavaForConditionalGeneration", "diffsynth.models.hunyuan_video_text_encoder", "hunyuan_video_text_encoder_2", "HunyuanVideoMLLMEncoder"),
     ("Step1Model", "diffsynth.models.stepvideo_text_encoder", "stepvideo_text_encoder_2", "STEP1TextEncoder"),
 ]
 patch_model_loader_configs = [
@@ -677,6 +678,25 @@ preset_models_on_modelscope = {
             "models/HunyuanVideo/transformers/mp_rank_00_model_states.pt"
         ],
     },
+    "HunyuanVideoI2V":{
+        "file_list": [
+            ("AI-ModelScope/clip-vit-large-patch14", "model.safetensors", "models/HunyuanVideoI2V/text_encoder"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model-00001-of-00004.safetensors", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model-00002-of-00004.safetensors", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model-00003-of-00004.safetensors", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model-00004-of-00004.safetensors", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "config.json", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/llava-llama-3-8b-v1_1-transformers", "model.safetensors.index.json", "models/HunyuanVideoI2V/text_encoder_2"),
+            ("AI-ModelScope/HunyuanVideo-I2V", "hunyuan-video-i2v-720p/vae/pytorch_model.pt", "models/HunyuanVideoI2V/vae"),
+            ("AI-ModelScope/HunyuanVideo-I2V", "hunyuan-video-i2v-720p/transformers/mp_rank_00_model_states.pt", "models/HunyuanVideoI2V/transformers")
+        ],
+        "load_path": [
+            "models/HunyuanVideoI2V/text_encoder/model.safetensors",
+            "models/HunyuanVideoI2V/text_encoder_2",
+            "models/HunyuanVideoI2V/vae/pytorch_model.pt",
+            "models/HunyuanVideoI2V/transformers/mp_rank_00_model_states.pt"
+        ],
+    },
     "HunyuanVideo-fp8":{
         "file_list": [
             ("AI-ModelScope/clip-vit-large-patch14", "model.safetensors", "models/HunyuanVideo/text_encoder"),
@@ -753,4 +773,5 @@ Preset_model_id: TypeAlias = Literal[
     "StableDiffusion3.5-medium",
     "HunyuanVideo",
     "HunyuanVideo-fp8",
+    "HunyuanVideoI2V",
 ]

diffsynth/models/hunyuan_video_dit.py

@@ -4,6 +4,7 @@ from .utils import init_weights_on_device
 from einops import rearrange, repeat
 from tqdm import tqdm
 from typing import Union, Tuple, List
+from .utils import hash_state_dict_keys
 
 
 def HunyuanVideoRope(latents):
@@ -281,7 +282,12 @@ class ModulateDiT(torch.nn.Module):
         return self.linear(self.act(x))
 
 
-def modulate(x, shift=None, scale=None):
+def modulate(x, shift=None, scale=None, tr_shift=None, tr_scale=None, tr_token=None):
+    if tr_shift is not None:
+        x_zero = x[:, :tr_token] * (1 + tr_scale.unsqueeze(1)) + tr_shift.unsqueeze(1)
+        x_orig = x[:, tr_token:] * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+        x = torch.concat((x_zero, x_orig), dim=1)
+        return x
     if scale is None and shift is None:
         return x
     elif shift is None:
@@ -385,6 +391,15 @@ def attention(q, k, v):
     return x
 
 
+def apply_gate(x, gate, tr_gate=None, tr_token=None):
+    if tr_gate is not None:
+        x_zero = x[:, :tr_token] * tr_gate.unsqueeze(1)
+        x_orig = x[:, tr_token:] * gate.unsqueeze(1)
+        return torch.concat((x_zero, x_orig), dim=1)
+    else:
+        return x * gate.unsqueeze(1)
+
+
 class MMDoubleStreamBlockComponent(torch.nn.Module):
     def __init__(self, hidden_size=3072, heads_num=24, mlp_width_ratio=4):
         super().__init__()
@@ -405,11 +420,17 @@ class MMDoubleStreamBlockComponent(torch.nn.Module):
             torch.nn.Linear(hidden_size * mlp_width_ratio, hidden_size)
         )
 
-    def forward(self, hidden_states, conditioning, freqs_cis=None):
+    def forward(self, hidden_states, conditioning, freqs_cis=None, token_replace_vec=None, tr_token=None):
         mod1_shift, mod1_scale, mod1_gate, mod2_shift, mod2_scale, mod2_gate = self.mod(conditioning).chunk(6, dim=-1)
+        if token_replace_vec is not None:
+            assert tr_token is not None
+            tr_mod1_shift, tr_mod1_scale, tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate = self.mod(token_replace_vec).chunk(6, dim=-1)
+        else:
+            tr_mod1_shift, tr_mod1_scale, tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate = None, None, None, None, None, None
 
         norm_hidden_states = self.norm1(hidden_states)
-        norm_hidden_states = modulate(norm_hidden_states, shift=mod1_shift, scale=mod1_scale)
+        norm_hidden_states = modulate(norm_hidden_states, shift=mod1_shift, scale=mod1_scale,
+                                      tr_shift=tr_mod1_shift, tr_scale=tr_mod1_scale, tr_token=tr_token)
         qkv = self.to_qkv(norm_hidden_states)
         q, k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.heads_num)
 
@@ -418,13 +439,17 @@ class MMDoubleStreamBlockComponent(torch.nn.Module):
 
         if freqs_cis is not None:
             q, k = apply_rotary_emb(q, k, freqs_cis, head_first=False)
+        return (q, k, v), (mod1_gate, mod2_shift, mod2_scale, mod2_gate), (tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate)
 
-        return (q, k, v), (mod1_gate, mod2_shift, mod2_scale, mod2_gate)
+    def process_ff(self, hidden_states, attn_output, mod, mod_tr=None, tr_token=None):
-    
-    def process_ff(self, hidden_states, attn_output, mod):
         mod1_gate, mod2_shift, mod2_scale, mod2_gate = mod
-        hidden_states = hidden_states + self.to_out(attn_output) * mod1_gate.unsqueeze(1)
+        if mod_tr is not None:
-        hidden_states = hidden_states + self.ff(modulate(self.norm2(hidden_states), shift=mod2_shift, scale=mod2_scale)) * mod2_gate.unsqueeze(1)
+            tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate = mod_tr
+        else:
+            tr_mod1_gate, tr_mod2_shift, tr_mod2_scale, tr_mod2_gate = None, None, None, None
+        hidden_states = hidden_states + apply_gate(self.to_out(attn_output), mod1_gate, tr_mod1_gate, tr_token)
+        x = self.ff(modulate(self.norm2(hidden_states), shift=mod2_shift, scale=mod2_scale, tr_shift=tr_mod2_shift, tr_scale=tr_mod2_scale, tr_token=tr_token))
+        hidden_states = hidden_states + apply_gate(x, mod2_gate, tr_mod2_gate, tr_token)
         return hidden_states
 
 
@@ -434,18 +459,18 @@ class MMDoubleStreamBlock(torch.nn.Module):
         self.component_a = MMDoubleStreamBlockComponent(hidden_size, heads_num, mlp_width_ratio)
         self.component_b = MMDoubleStreamBlockComponent(hidden_size, heads_num, mlp_width_ratio)
 
-    def forward(self, hidden_states_a, hidden_states_b, conditioning, freqs_cis):
+    def forward(self, hidden_states_a, hidden_states_b, conditioning, freqs_cis, token_replace_vec=None, tr_token=None, split_token=71):
-        (q_a, k_a, v_a), mod_a = self.component_a(hidden_states_a, conditioning, freqs_cis)
+        (q_a, k_a, v_a), mod_a, mod_tr = self.component_a(hidden_states_a, conditioning, freqs_cis, token_replace_vec, tr_token)
-        (q_b, k_b, v_b), mod_b = self.component_b(hidden_states_b, conditioning, freqs_cis=None)
+        (q_b, k_b, v_b), mod_b, _ = self.component_b(hidden_states_b, conditioning, freqs_cis=None)
 
-        q_a, q_b = torch.concat([q_a, q_b[:, :71]], dim=1), q_b[:, 71:].contiguous()
+        q_a, q_b = torch.concat([q_a, q_b[:, :split_token]], dim=1), q_b[:, split_token:].contiguous()
-        k_a, k_b = torch.concat([k_a, k_b[:, :71]], dim=1), k_b[:, 71:].contiguous()
+        k_a, k_b = torch.concat([k_a, k_b[:, :split_token]], dim=1), k_b[:, split_token:].contiguous()
-        v_a, v_b = torch.concat([v_a, v_b[:, :71]], dim=1), v_b[:, 71:].contiguous()
+        v_a, v_b = torch.concat([v_a, v_b[:, :split_token]], dim=1), v_b[:, split_token:].contiguous()
         attn_output_a = attention(q_a, k_a, v_a)
         attn_output_b = attention(q_b, k_b, v_b)
-        attn_output_a, attn_output_b = attn_output_a[:, :-71].contiguous(), torch.concat([attn_output_a[:, -71:], attn_output_b], dim=1)
+        attn_output_a, attn_output_b = attn_output_a[:, :-split_token].contiguous(), torch.concat([attn_output_a[:, -split_token:], attn_output_b], dim=1)
 
-        hidden_states_a = self.component_a.process_ff(hidden_states_a, attn_output_a, mod_a)
+        hidden_states_a = self.component_a.process_ff(hidden_states_a, attn_output_a, mod_a, mod_tr, tr_token)
         hidden_states_b = self.component_b.process_ff(hidden_states_b, attn_output_b, mod_b)
         return hidden_states_a, hidden_states_b
 
@@ -509,11 +534,17 @@ class MMSingleStreamBlock(torch.nn.Module):
             torch.nn.Linear(hidden_size * mlp_width_ratio, hidden_size, bias=False)
         )
 
-    def forward(self, hidden_states, conditioning, freqs_cis=None, txt_len=256):
+    def forward(self, hidden_states, conditioning, freqs_cis=None, txt_len=256, token_replace_vec=None, tr_token=None, split_token=71):
         mod_shift, mod_scale, mod_gate = self.mod(conditioning).chunk(3, dim=-1)
+        if token_replace_vec is not None:
+            assert tr_token is not None
+            tr_mod_shift, tr_mod_scale, tr_mod_gate = self.mod(token_replace_vec).chunk(3, dim=-1)
+        else:
+            tr_mod_shift, tr_mod_scale, tr_mod_gate = None, None, None
 
         norm_hidden_states = self.norm(hidden_states)
-        norm_hidden_states = modulate(norm_hidden_states, shift=mod_shift, scale=mod_scale)
+        norm_hidden_states = modulate(norm_hidden_states, shift=mod_shift, scale=mod_scale,
+                                      tr_shift=tr_mod_shift, tr_scale=tr_mod_scale, tr_token=tr_token)
         qkv = self.to_qkv(norm_hidden_states)
 
         q, k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.heads_num)
@@ -525,16 +556,17 @@ class MMSingleStreamBlock(torch.nn.Module):
         k_a, k_b = k[:, :-txt_len, :, :], k[:, -txt_len:, :, :]
         q_a, k_a = apply_rotary_emb(q_a, k_a, freqs_cis, head_first=False)
 
-        q_a, q_b = torch.concat([q_a, q_b[:, :71]], dim=1), q_b[:, 71:].contiguous()
+        v_len = txt_len - split_token
-        k_a, k_b = torch.concat([k_a, k_b[:, :71]], dim=1), k_b[:, 71:].contiguous()
+        q_a, q_b = torch.concat([q_a, q_b[:, :split_token]], dim=1), q_b[:, split_token:].contiguous()
-        v_a, v_b = v[:, :-185].contiguous(), v[:, -185:].contiguous()
+        k_a, k_b = torch.concat([k_a, k_b[:, :split_token]], dim=1), k_b[:, split_token:].contiguous()
+        v_a, v_b = v[:, :-v_len].contiguous(), v[:, -v_len:].contiguous()
 
         attn_output_a = attention(q_a, k_a, v_a)
         attn_output_b = attention(q_b, k_b, v_b)
         attn_output = torch.concat([attn_output_a, attn_output_b], dim=1)
 
-        hidden_states = hidden_states + self.to_out(attn_output) * mod_gate.unsqueeze(1)
+        hidden_states = hidden_states + apply_gate(self.to_out(attn_output), mod_gate, tr_mod_gate, tr_token)
-        hidden_states = hidden_states + self.ff(norm_hidden_states) * mod_gate.unsqueeze(1)
+        hidden_states = hidden_states + apply_gate(self.ff(norm_hidden_states), mod_gate, tr_mod_gate, tr_token)
         return hidden_states
 
 
@@ -555,7 +587,7 @@ class FinalLayer(torch.nn.Module):
 
 
 class HunyuanVideoDiT(torch.nn.Module):
-    def __init__(self, in_channels=16, hidden_size=3072, text_dim=4096, num_double_blocks=20, num_single_blocks=40):
+    def __init__(self, in_channels=16, hidden_size=3072, text_dim=4096, num_double_blocks=20, num_single_blocks=40, guidance_embed=True):
         super().__init__()
         self.img_in = PatchEmbed(in_channels=in_channels, embed_dim=hidden_size)
         self.txt_in = SingleTokenRefiner(in_channels=text_dim, hidden_size=hidden_size)
@@ -565,7 +597,7 @@ class HunyuanVideoDiT(torch.nn.Module):
             torch.nn.SiLU(),
             torch.nn.Linear(hidden_size, hidden_size)
         )
-        self.guidance_in = TimestepEmbeddings(256, hidden_size, computation_device="cpu")
+        self.guidance_in = TimestepEmbeddings(256, hidden_size, computation_device="cpu") if guidance_embed else None
         self.double_blocks = torch.nn.ModuleList([MMDoubleStreamBlock(hidden_size) for _ in range(num_double_blocks)])
         self.single_blocks = torch.nn.ModuleList([MMSingleStreamBlock(hidden_size) for _ in range(num_single_blocks)])
         self.final_layer = FinalLayer(hidden_size)
@@ -610,7 +642,9 @@ class HunyuanVideoDiT(torch.nn.Module):
     ):
         B, C, T, H, W = x.shape
 
-        vec = self.time_in(t, dtype=torch.float32) + self.vector_in(pooled_prompt_emb) + self.guidance_in(guidance * 1000, dtype=torch.float32)
+        vec = self.time_in(t, dtype=torch.float32) + self.vector_in(pooled_prompt_emb)
+        if self.guidance_in is not None:
+            vec += self.guidance_in(guidance * 1000, dtype=torch.float32)
         img = self.img_in(x)
         txt = self.txt_in(prompt_emb, t, text_mask)
 
@@ -777,12 +811,12 @@ class HunyuanVideoDiT(torch.nn.Module):
         return HunyuanVideoDiTStateDictConverter()
 
 
-
 class HunyuanVideoDiTStateDictConverter:
     def __init__(self):
         pass
 
     def from_civitai(self, state_dict):
+        origin_hash_key = hash_state_dict_keys(state_dict, with_shape=True)
         if "module" in state_dict:
             state_dict = state_dict["module"]
         direct_dict = {
@@ -882,4 +916,5 @@ class HunyuanVideoDiTStateDictConverter:
                 state_dict_[name_] = param
             else:
                 pass
+
         return state_dict_

diffsynth/models/hunyuan_video_text_encoder.py

@@ -1,24 +1,18 @@
-from transformers import LlamaModel, LlamaConfig, DynamicCache
+from transformers import LlamaModel, LlamaConfig, DynamicCache, LlavaForConditionalGeneration
 from copy import deepcopy
 import torch
 
 
 class HunyuanVideoLLMEncoder(LlamaModel):
+
     def __init__(self, config: LlamaConfig):
         super().__init__(config)
         self.auto_offload = False
 
-
     def enable_auto_offload(self, **kwargs):
         self.auto_offload = True
 
-
+    def forward(self, input_ids, attention_mask, hidden_state_skip_layer=2):
-    def forward(
-        self,
-        input_ids,
-        attention_mask,
-        hidden_state_skip_layer=2
-    ):
         embed_tokens = deepcopy(self.embed_tokens).to(input_ids.device) if self.auto_offload else self.embed_tokens
         inputs_embeds = embed_tokens(input_ids)
 
@@ -53,3 +47,22 @@ class HunyuanVideoLLMEncoder(LlamaModel):
                 break
 
         return hidden_states
+
+
+class HunyuanVideoMLLMEncoder(LlavaForConditionalGeneration):
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.auto_offload = False
+
+    def enable_auto_offload(self, **kwargs):
+        self.auto_offload = True
+
+    # TODO: implement the low VRAM inference for MLLM.
+    def forward(self, input_ids, pixel_values, attention_mask, hidden_state_skip_layer=2):
+        outputs = super().forward(input_ids=input_ids,
+                                  attention_mask=attention_mask,
+                                  output_hidden_states=True,
+                                  pixel_values=pixel_values)
+        hidden_state = outputs.hidden_states[-(hidden_state_skip_layer + 1)]
+        return hidden_state

diffsynth/pipelines/hunyuan_video.py

@@ -5,13 +5,13 @@ from ..schedulers.flow_match import FlowMatchScheduler
 from .base import BasePipeline
 from ..prompters import HunyuanVideoPrompter
 import torch
+import torchvision.transforms as transforms
 from einops import rearrange
 import numpy as np
 from PIL import Image
 from tqdm import tqdm
 
 
-
 class HunyuanVideoPipeline(BasePipeline):
 
     def __init__(self, device="cuda", torch_dtype=torch.float16):
@@ -53,10 +53,58 @@ class HunyuanVideoPipeline(BasePipeline):
             pipe.enable_vram_management()
         return pipe
 
+    def generate_crop_size_list(self, base_size=256, patch_size=32, max_ratio=4.0):
+        num_patches = round((base_size / patch_size)**2)
+        assert max_ratio >= 1.0
+        crop_size_list = []
+        wp, hp = num_patches, 1
+        while wp > 0:
+            if max(wp, hp) / min(wp, hp) <= max_ratio:
+                crop_size_list.append((wp * patch_size, hp * patch_size))
+            if (hp + 1) * wp <= num_patches:
+                hp += 1
+            else:
+                wp -= 1
+        return crop_size_list
 
-    def encode_prompt(self, prompt, positive=True, clip_sequence_length=77, llm_sequence_length=256):
+
+    def get_closest_ratio(self, height: float, width: float, ratios: list, buckets: list):
+        aspect_ratio = float(height) / float(width)
+        closest_ratio_id = np.abs(ratios - aspect_ratio).argmin()
+        closest_ratio = min(ratios, key=lambda ratio: abs(float(ratio) - aspect_ratio))
+        return buckets[closest_ratio_id], float(closest_ratio)
+
+
+    def prepare_vae_images_inputs(self, semantic_images, i2v_resolution="720p"):
+        if i2v_resolution == "720p":
+            bucket_hw_base_size = 960
+        elif i2v_resolution == "540p":
+            bucket_hw_base_size = 720
+        elif i2v_resolution == "360p":
+            bucket_hw_base_size = 480
+        else:
+            raise ValueError(f"i2v_resolution: {i2v_resolution} must be in [360p, 540p, 720p]")
+        origin_size = semantic_images[0].size
+
+        crop_size_list = self.generate_crop_size_list(bucket_hw_base_size, 32)
+        aspect_ratios = np.array([round(float(h) / float(w), 5) for h, w in crop_size_list])
+        closest_size, closest_ratio = self.get_closest_ratio(origin_size[1], origin_size[0], aspect_ratios, crop_size_list)
+        ref_image_transform = transforms.Compose([
+            transforms.Resize(closest_size),
+            transforms.CenterCrop(closest_size),
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5])
+        ])
+
+        semantic_image_pixel_values = [ref_image_transform(semantic_image) for semantic_image in semantic_images]
+        semantic_image_pixel_values = torch.cat(semantic_image_pixel_values).unsqueeze(0).unsqueeze(2).to(self.device)
+        target_height, target_width = closest_size
+        return semantic_image_pixel_values, target_height, target_width
+
+
+    def encode_prompt(self, prompt, positive=True, clip_sequence_length=77, llm_sequence_length=256, input_images=None):
         prompt_emb, pooled_prompt_emb, text_mask = self.prompter.encode_prompt(
-            prompt, device=self.device, positive=positive, clip_sequence_length=clip_sequence_length, llm_sequence_length=llm_sequence_length
+            prompt, device=self.device, positive=positive, clip_sequence_length=clip_sequence_length, llm_sequence_length=llm_sequence_length, images=input_images
         )
         return {"prompt_emb": prompt_emb, "pooled_prompt_emb": pooled_prompt_emb, "text_mask": text_mask}
 
@@ -87,6 +135,9 @@ class HunyuanVideoPipeline(BasePipeline):
         prompt,
         negative_prompt="",
         input_video=None,
+        input_images=None,
+        i2v_resolution="720p",
+        i2v_stability=True,
         denoising_strength=1.0,
         seed=None,
         rand_device=None,
@@ -109,6 +160,13 @@ class HunyuanVideoPipeline(BasePipeline):
         # Scheduler
         self.scheduler.set_timesteps(num_inference_steps, denoising_strength)
 
+        # encoder input images
+        if input_images is not None:
+            self.load_models_to_device(['vae_encoder'])
+            image_pixel_values, height, width = self.prepare_vae_images_inputs(input_images, i2v_resolution=i2v_resolution)
+            with torch.autocast(device_type=self.device, dtype=torch.float16, enabled=True):
+                image_latents = self.vae_encoder(image_pixel_values)
+
         # Initialize noise
         rand_device = self.device if rand_device is None else rand_device
         noise = self.generate_noise((1, 16, (num_frames - 1) // 4 + 1, height//8, width//8), seed=seed, device=rand_device, dtype=self.torch_dtype).to(self.device)
@@ -118,12 +176,18 @@ class HunyuanVideoPipeline(BasePipeline):
             input_video = torch.stack(input_video, dim=2)
             latents = self.encode_video(input_video, **tiler_kwargs).to(dtype=self.torch_dtype, device=self.device)
             latents = self.scheduler.add_noise(latents, noise, timestep=self.scheduler.timesteps[0])
+        elif input_images is not None and i2v_stability:
+            noise = self.generate_noise((1, 16, (num_frames - 1) // 4 + 1, height//8, width//8), seed=seed, device=rand_device, dtype=image_latents.dtype).to(self.device)
+            t = torch.tensor([0.999]).to(device=self.device)
+            latents = noise * t + image_latents.repeat(1, 1, (num_frames - 1) // 4 + 1, 1, 1) * (1 - t)
+            latents = latents.to(dtype=image_latents.dtype)
         else:
             latents = noise
 
         # Encode prompts
-        self.load_models_to_device(["text_encoder_1"] if self.vram_management else ["text_encoder_1", "text_encoder_2"])
+        # current mllm does not support vram_management
-        prompt_emb_posi = self.encode_prompt(prompt, positive=True)
+        self.load_models_to_device(["text_encoder_1"] if self.vram_management and input_images is None else ["text_encoder_1", "text_encoder_2"])
+        prompt_emb_posi = self.encode_prompt(prompt, positive=True, input_images=input_images)
         if cfg_scale != 1.0:
             prompt_emb_nega = self.encode_prompt(negative_prompt, positive=False)
 
@@ -139,11 +203,16 @@ class HunyuanVideoPipeline(BasePipeline):
             timestep = timestep.unsqueeze(0).to(self.device)
             print(f"Step {progress_id + 1} / {len(self.scheduler.timesteps)}")
 
+            forward_func = lets_dance_hunyuan_video
+            if input_images is not None:
+                latents = torch.concat([image_latents, latents[:, :, 1:, :, :]], dim=2)
+                forward_func = lets_dance_hunyuan_video_i2v
+
             # Inference
             with torch.autocast(device_type=self.device, dtype=self.torch_dtype):
-                noise_pred_posi = lets_dance_hunyuan_video(self.dit, latents, timestep, **prompt_emb_posi, **extra_input, **tea_cache_kwargs)
+                noise_pred_posi = forward_func(self.dit, latents, timestep, **prompt_emb_posi, **extra_input, **tea_cache_kwargs)
                 if cfg_scale != 1.0:
-                    noise_pred_nega = lets_dance_hunyuan_video(self.dit, latents, timestep, **prompt_emb_nega, **extra_input)
+                    noise_pred_nega = forward_func(self.dit, latents, timestep, **prompt_emb_nega, **extra_input)
                     noise_pred = noise_pred_nega + cfg_scale * (noise_pred_posi - noise_pred_nega)
                 else:
                     noise_pred = noise_pred_posi
@@ -163,7 +232,11 @@ class HunyuanVideoPipeline(BasePipeline):
                 self.load_models_to_device([] if self.vram_management else ["dit"])
 
             # Scheduler
-            latents = self.scheduler.step(noise_pred, self.scheduler.timesteps[progress_id], latents)
+            if input_images is not None:
+                latents = self.scheduler.step(noise_pred[:, :, 1:, :, :], self.scheduler.timesteps[progress_id], latents[:, :, 1:, :, :])
+                latents = torch.concat([image_latents, latents], dim=2)
+            else:
+                latents = self.scheduler.step(noise_pred, self.scheduler.timesteps[progress_id], latents)
 
         # Decode
         self.load_models_to_device(['vae_decoder'])
@@ -250,13 +323,70 @@ def lets_dance_hunyuan_video(
         print("TeaCache skip forward.")
         img = tea_cache.update(img)
     else:
+        split_token = int(text_mask.sum(dim=1))
+        txt_len = int(txt.shape[1])
         for block in tqdm(dit.double_blocks, desc="Double stream blocks"):
-            img, txt = block(img, txt, vec, (freqs_cos, freqs_sin))
+            img, txt = block(img, txt, vec, (freqs_cos, freqs_sin), split_token=split_token)
 
         x = torch.concat([img, txt], dim=1)
         for block in tqdm(dit.single_blocks, desc="Single stream blocks"):
-            x = block(x, vec, (freqs_cos, freqs_sin))
+            x = block(x, vec, (freqs_cos, freqs_sin), txt_len=txt_len, split_token=split_token)
-        img = x[:, :-256]
+        img = x[:, :-txt_len]
+
+        if tea_cache is not None:
+            tea_cache.store(img)
+    img = dit.final_layer(img, vec)
+    img = dit.unpatchify(img, T=T//1, H=H//2, W=W//2)
+    return img
+
+
+def lets_dance_hunyuan_video_i2v(
+    dit: HunyuanVideoDiT,
+    x: torch.Tensor,
+    t: torch.Tensor,
+    prompt_emb: torch.Tensor = None,
+    text_mask: torch.Tensor = None,
+    pooled_prompt_emb: torch.Tensor = None,
+    freqs_cos: torch.Tensor = None,
+    freqs_sin: torch.Tensor = None,
+    guidance: torch.Tensor = None,
+    tea_cache: TeaCache = None,
+    **kwargs
+):
+    B, C, T, H, W = x.shape
+    # Uncomment below to keep same as official implementation
+    # guidance = guidance.to(dtype=torch.float32).to(torch.bfloat16)
+    vec = dit.time_in(t, dtype=torch.bfloat16)
+    vec_2 = dit.vector_in(pooled_prompt_emb)
+    vec = vec + vec_2
+    vec = vec + dit.guidance_in(guidance * 1000., dtype=torch.bfloat16)
+
+    token_replace_vec = dit.time_in(torch.zeros_like(t), dtype=torch.bfloat16)
+    tr_token = (H // 2) * (W // 2)
+    token_replace_vec = token_replace_vec + vec_2
+
+    img = dit.img_in(x)
+    txt = dit.txt_in(prompt_emb, t, text_mask)
+
+    # TeaCache
+    if tea_cache is not None:
+        tea_cache_update = tea_cache.check(dit, img, vec)
+    else:
+        tea_cache_update = False
+
+    if tea_cache_update:
+        print("TeaCache skip forward.")
+        img = tea_cache.update(img)
+    else:
+        split_token = int(text_mask.sum(dim=1))
+        txt_len = int(txt.shape[1])
+        for block in tqdm(dit.double_blocks, desc="Double stream blocks"):
+            img, txt = block(img, txt, vec, (freqs_cos, freqs_sin), token_replace_vec, tr_token, split_token)
+
+        x = torch.concat([img, txt], dim=1)
+        for block in tqdm(dit.single_blocks, desc="Single stream blocks"):
+            x = block(x, vec, (freqs_cos, freqs_sin), txt_len, token_replace_vec, tr_token, split_token)
+        img = x[:, :-txt_len]
 
         if tea_cache is not None:
             tea_cache.store(img)

diffsynth/prompters/hunyuan_video_prompter.py

@@ -1,8 +1,9 @@
 from .base_prompter import BasePrompter
 from ..models.sd3_text_encoder import SD3TextEncoder1
-from ..models.hunyuan_video_text_encoder import HunyuanVideoLLMEncoder
+from ..models.hunyuan_video_text_encoder import HunyuanVideoLLMEncoder, HunyuanVideoMLLMEncoder
-from transformers import CLIPTokenizer, LlamaTokenizerFast
+from transformers import CLIPTokenizer, LlamaTokenizerFast, CLIPImageProcessor
 import os, torch
+from typing import Union
 
 PROMPT_TEMPLATE_ENCODE = (
     "<|start_header_id|>system<|end_header_id|>\n\nDescribe the image by detailing the color, shape, size, texture, "
@@ -18,6 +19,24 @@ PROMPT_TEMPLATE_ENCODE_VIDEO = (
     "5. camera angles, movements, and transitions used in the video:<|eot_id|>"
     "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>")
 
+PROMPT_TEMPLATE_ENCODE_I2V = (
+    "<|start_header_id|>system<|end_header_id|>\n\n<image>\nDescribe the image by detailing the color, shape, size, texture, "
+    "quantity, text, spatial relationships of the objects and background:<|eot_id|>"
+    "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>"
+    "<|start_header_id|>assistant<|end_header_id|>\n\n"
+)
+
+PROMPT_TEMPLATE_ENCODE_VIDEO_I2V = (
+    "<|start_header_id|>system<|end_header_id|>\n\n<image>\nDescribe the video by detailing the following aspects according to the reference image: "
+    "1. The main content and theme of the video."
+    "2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects."
+    "3. Actions, events, behaviors temporal relationships, physical movement changes of the objects."
+    "4. background environment, light, style and atmosphere."
+    "5. camera angles, movements, and transitions used in the video:<|eot_id|>\n\n"
+    "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>"
+    "<|start_header_id|>assistant<|end_header_id|>\n\n"
+)
+
 PROMPT_TEMPLATE = {
     "dit-llm-encode": {
         "template": PROMPT_TEMPLATE_ENCODE,
@@ -27,6 +46,22 @@ PROMPT_TEMPLATE = {
         "template": PROMPT_TEMPLATE_ENCODE_VIDEO,
         "crop_start": 95,
     },
+    "dit-llm-encode-i2v": {
+        "template": PROMPT_TEMPLATE_ENCODE_I2V,
+        "crop_start": 36,
+        "image_emb_start": 5,
+        "image_emb_end": 581,
+        "image_emb_len": 576,
+        "double_return_token_id": 271
+    },
+    "dit-llm-encode-video-i2v": {
+        "template": PROMPT_TEMPLATE_ENCODE_VIDEO_I2V,
+        "crop_start": 103,
+        "image_emb_start": 5,
+        "image_emb_end": 581,
+        "image_emb_len": 576,
+        "double_return_token_id": 271
+    },
 }
 
 NEGATIVE_PROMPT = "Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion"
@@ -56,9 +91,20 @@ class HunyuanVideoPrompter(BasePrompter):
         self.prompt_template = PROMPT_TEMPLATE['dit-llm-encode']
         self.prompt_template_video = PROMPT_TEMPLATE['dit-llm-encode-video']
 
-    def fetch_models(self, text_encoder_1: SD3TextEncoder1 = None, text_encoder_2: HunyuanVideoLLMEncoder = None):
+    def fetch_models(self,
+                     text_encoder_1: SD3TextEncoder1 = None,
+                     text_encoder_2: Union[HunyuanVideoLLMEncoder, HunyuanVideoMLLMEncoder] = None):
         self.text_encoder_1 = text_encoder_1
         self.text_encoder_2 = text_encoder_2
+        if isinstance(text_encoder_2, HunyuanVideoMLLMEncoder):
+            # processor
+            # TODO: may need to replace processor with local implementation
+            base_path = os.path.dirname(os.path.dirname(__file__))
+            tokenizer_2_path = os.path.join(base_path, "tokenizer_configs/hunyuan_video/tokenizer_2")
+            self.processor = CLIPImageProcessor.from_pretrained(tokenizer_2_path)
+            # template
+            self.prompt_template = PROMPT_TEMPLATE['dit-llm-encode-i2v']
+            self.prompt_template_video = PROMPT_TEMPLATE['dit-llm-encode-video-i2v']
 
     def apply_text_to_template(self, text, template):
         assert isinstance(template, str)
@@ -107,8 +153,89 @@ class HunyuanVideoPrompter(BasePrompter):
 
         return last_hidden_state, attention_mask
 
+    def encode_prompt_using_mllm(self,
+                                prompt,
+                                images,
+                                max_length,
+                                device,
+                                crop_start,
+                                hidden_state_skip_layer=2,
+                                use_attention_mask=True,
+                                image_embed_interleave=4):
+        image_outputs = self.processor(images, return_tensors="pt")["pixel_values"].to(device)
+        max_length += crop_start
+        inputs = self.tokenizer_2(prompt,
+                                  return_tensors="pt",
+                                  padding="max_length",
+                                  max_length=max_length,
+                                  truncation=True)
+        input_ids = inputs.input_ids.to(device)
+        attention_mask = inputs.attention_mask.to(device)
+        last_hidden_state = self.text_encoder_2(input_ids=input_ids,
+                                                attention_mask=attention_mask,
+                                                hidden_state_skip_layer=hidden_state_skip_layer,
+                                                pixel_values=image_outputs)
+
+        text_crop_start = (crop_start - 1 + self.prompt_template_video.get("image_emb_len", 576))
+        image_crop_start = self.prompt_template_video.get("image_emb_start", 5)
+        image_crop_end = self.prompt_template_video.get("image_emb_end", 581)
+        batch_indices, last_double_return_token_indices = torch.where(
+            input_ids == self.prompt_template_video.get("double_return_token_id", 271))
+        if last_double_return_token_indices.shape[0] == 3:
+            # in case the prompt is too long
+            last_double_return_token_indices = torch.cat((
+                last_double_return_token_indices,
+                torch.tensor([input_ids.shape[-1]]),
+            ))
+            batch_indices = torch.cat((batch_indices, torch.tensor([0])))
+        last_double_return_token_indices = (last_double_return_token_indices.reshape(input_ids.shape[0], -1)[:, -1])
+        batch_indices = batch_indices.reshape(input_ids.shape[0], -1)[:, -1]
+        assistant_crop_start = (last_double_return_token_indices - 1 + self.prompt_template_video.get("image_emb_len", 576) - 4)
+        assistant_crop_end = (last_double_return_token_indices - 1 + self.prompt_template_video.get("image_emb_len", 576))
+        attention_mask_assistant_crop_start = (last_double_return_token_indices - 4)
+        attention_mask_assistant_crop_end = last_double_return_token_indices
+        text_last_hidden_state = []
+        text_attention_mask = []
+        image_last_hidden_state = []
+        image_attention_mask = []
+        for i in range(input_ids.shape[0]):
+            text_last_hidden_state.append(
+                torch.cat([
+                    last_hidden_state[i, text_crop_start:assistant_crop_start[i].item()],
+                    last_hidden_state[i, assistant_crop_end[i].item():],
+                ]))
+            text_attention_mask.append(
+                torch.cat([
+                    attention_mask[
+                        i,
+                        crop_start:attention_mask_assistant_crop_start[i].item(),
+                    ],
+                    attention_mask[i, attention_mask_assistant_crop_end[i].item():],
+                ]) if use_attention_mask else None)
+            image_last_hidden_state.append(last_hidden_state[i, image_crop_start:image_crop_end])
+            image_attention_mask.append(
+                torch.ones(image_last_hidden_state[-1].shape[0]).to(last_hidden_state.device).
+                to(attention_mask.dtype) if use_attention_mask else None)
+
+        text_last_hidden_state = torch.stack(text_last_hidden_state)
+        text_attention_mask = torch.stack(text_attention_mask)
+        image_last_hidden_state = torch.stack(image_last_hidden_state)
+        image_attention_mask = torch.stack(image_attention_mask)
+
+        image_last_hidden_state = image_last_hidden_state[:, ::image_embed_interleave, :]
+        image_attention_mask = image_attention_mask[:, ::image_embed_interleave]
+
+        assert (text_last_hidden_state.shape[0] == text_attention_mask.shape[0] and
+                image_last_hidden_state.shape[0] == image_attention_mask.shape[0])
+
+        last_hidden_state = torch.cat([image_last_hidden_state, text_last_hidden_state], dim=1)
+        attention_mask = torch.cat([image_attention_mask, text_attention_mask], dim=1)
+
+        return last_hidden_state, attention_mask
+
     def encode_prompt(self,
                       prompt,
+                      images=None,
                       positive=True,
                       device="cuda",
                       clip_sequence_length=77,
@@ -116,7 +243,8 @@ class HunyuanVideoPrompter(BasePrompter):
                       data_type='video',
                       use_template=True,
                       hidden_state_skip_layer=2,
-                      use_attention_mask=True):
+                      use_attention_mask=True,
+                      image_embed_interleave=4):
 
         prompt = self.process_prompt(prompt, positive=positive)
 
@@ -136,8 +264,12 @@ class HunyuanVideoPrompter(BasePrompter):
         pooled_prompt_emb = self.encode_prompt_using_clip(prompt, clip_sequence_length, device)
 
         # LLM
-        prompt_emb, attention_mask = self.encode_prompt_using_llm(
+        if images is None:
-            prompt_formated, llm_sequence_length, device, crop_start,
+            prompt_emb, attention_mask = self.encode_prompt_using_llm(prompt_formated, llm_sequence_length, device, crop_start,
-            hidden_state_skip_layer, use_attention_mask)
+                                                                      hidden_state_skip_layer, use_attention_mask)
+        else:
+            prompt_emb, attention_mask = self.encode_prompt_using_mllm(prompt_formated, images, llm_sequence_length, device,
+                                                                       crop_start, hidden_state_skip_layer, use_attention_mask,
+                                                                       image_embed_interleave)
 
         return prompt_emb, pooled_prompt_emb, attention_mask

diffsynth/tokenizer_configs/hunyuan_video/tokenizer_2/preprocessor_config.json

@@ -0,0 +1,45 @@
+{
+  "_valid_processor_keys": [
+    "images",
+    "do_resize",
+    "size",
+    "resample",
+    "do_center_crop",
+    "crop_size",
+    "do_rescale",
+    "rescale_factor",
+    "do_normalize",
+    "image_mean",
+    "image_std",
+    "do_convert_rgb",
+    "return_tensors",
+    "data_format",
+    "input_data_format"
+  ],
+  "crop_size": {
+    "height": 336,
+    "width": 336
+  },
+  "do_center_crop": true,
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_processor_type": "CLIPImageProcessor",
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "processor_class": "LlavaProcessor",
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "shortest_edge": 336
+  }
+}

examples/HunyuanVideo/README.md

@@ -8,6 +8,12 @@
 |24G|[hunyuanvideo_24G.py](hunyuanvideo_24G.py)|129|720*1280|The video is consistent with the original implementation, but it requires 5%~10% more time than [hunyuanvideo_80G.py](hunyuanvideo_80G.py)|
 |6G|[hunyuanvideo_6G.py](hunyuanvideo_6G.py)|129|512*384|The base model doesn't support low resolutions. We recommend users to use some LoRA ([example](https://civitai.com/models/1032126/walking-animation-hunyuan-video)) trained using low resolutions.|
 
+[HunyuanVideo-I2V](https://github.com/Tencent/HunyuanVideo-I2V) is the image-to-video generation version of HunyuanVideo. We also provide advanced VRAM management for this model.
+|VRAM required|Example script|Frames|Resolution|Note|
+|-|-|-|-|-|
+|80G|[hunyuanvideo_i2v_80G.py](hunyuanvideo_i2v_80G.py)|129|720p|No VRAM management.|
+|24G|[hunyuanvideo_i2v_24G.py](hunyuanvideo_i2v_24G.py)|129|720p|The video is consistent with the original implementation, but it requires 5%~10% more time than [hunyuanvideo_80G.py](hunyuanvideo_80G.py)|
+
 ## Gallery
 
 Video generated by [hunyuanvideo_80G.py](hunyuanvideo_80G.py) and [hunyuanvideo_24G.py](hunyuanvideo_24G.py):
@@ -21,3 +27,7 @@ https://github.com/user-attachments/assets/2997f107-d02d-4ecb-89bb-5ce1a7f93817
 Video to video generated by [hunyuanvideo_v2v_6G.py](./hunyuanvideo_v2v_6G.py) using [this LoRA](https://civitai.com/models/1032126/walking-animation-hunyuan-video):
 
 https://github.com/user-attachments/assets/4b89e52e-ce42-434e-aa57-08f09dfa2b10
+
+Video generated by [hunyuanvideo_i2v_80G.py](hunyuanvideo_i2v_80G.py) and [hunyuanvideo_i2v_24G.py](hunyuanvideo_i2v_24G.py):
+
+https://github.com/user-attachments/assets/494f252a-c9af-440d-84ba-a8ddcdcc538a

examples/HunyuanVideo/hunyuanvideo_i2v_24G.py

@@ -0,0 +1,43 @@
+import torch
+from diffsynth import ModelManager, HunyuanVideoPipeline, download_models, save_video
+from modelscope import dataset_snapshot_download
+from PIL import Image
+
+
+download_models(["HunyuanVideoI2V"])
+model_manager = ModelManager()
+
+# The DiT model is loaded in bfloat16.
+model_manager.load_models(
+    [
+        "models/HunyuanVideoI2V/transformers/mp_rank_00_model_states.pt"
+    ],
+    torch_dtype=torch.bfloat16,
+    device="cpu"
+)
+
+# The other modules are loaded in float16.
+model_manager.load_models(
+    [
+        "models/HunyuanVideoI2V/text_encoder/model.safetensors",
+        "models/HunyuanVideoI2V/text_encoder_2",
+        'models/HunyuanVideoI2V/vae/pytorch_model.pt'
+    ],
+    torch_dtype=torch.float16,
+    device="cpu"
+)
+# The computation device is "cuda".
+pipe = HunyuanVideoPipeline.from_model_manager(model_manager,
+                                               torch_dtype=torch.bfloat16,
+                                               device="cuda",
+                                               enable_vram_management=True)
+
+dataset_snapshot_download(dataset_id="DiffSynth-Studio/examples_in_diffsynth",
+                          local_dir="./",
+                          allow_file_pattern=f"data/examples/hunyuanvideo/*")
+
+i2v_resolution = "720p"
+prompt = "An Asian man with short hair in black tactical uniform and white clothes waves a firework stick."
+images = [Image.open("data/examples/hunyuanvideo/0.jpg").convert('RGB')]
+video = pipe(prompt, input_images=images, num_inference_steps=50, seed=0, i2v_resolution=i2v_resolution)
+save_video(video, f"video_{i2v_resolution}_low_vram.mp4", fps=30, quality=6)

examples/HunyuanVideo/hunyuanvideo_i2v_80G.py

@@ -0,0 +1,45 @@
+import torch
+from diffsynth import ModelManager, HunyuanVideoPipeline, download_models, save_video
+from modelscope import dataset_snapshot_download
+from PIL import Image
+
+
+download_models(["HunyuanVideoI2V"])
+model_manager = ModelManager()
+
+# The DiT model is loaded in bfloat16.
+model_manager.load_models(
+    [
+        "models/HunyuanVideoI2V/transformers/mp_rank_00_model_states.pt"
+    ],
+    torch_dtype=torch.bfloat16,
+    device="cuda"
+)
+
+# The other modules are loaded in float16.
+model_manager.load_models(
+    [
+        "models/HunyuanVideoI2V/text_encoder/model.safetensors",
+        "models/HunyuanVideoI2V/text_encoder_2",
+        'models/HunyuanVideoI2V/vae/pytorch_model.pt'
+    ],
+    torch_dtype=torch.float16,
+    device="cuda"
+)
+# The computation device is "cuda".
+pipe = HunyuanVideoPipeline.from_model_manager(model_manager,
+                                               torch_dtype=torch.bfloat16,
+                                               device="cuda",
+                                               enable_vram_management=False)
+# Although you have enough VRAM, we still recommend you to enable offload.
+pipe.enable_cpu_offload()
+
+dataset_snapshot_download(dataset_id="DiffSynth-Studio/examples_in_diffsynth",
+                          local_dir="./",
+                          allow_file_pattern=f"data/examples/hunyuanvideo/*")
+
+i2v_resolution = "720p"
+prompt = "An Asian man with short hair in black tactical uniform and white clothes waves a firework stick."
+images = [Image.open("data/examples/hunyuanvideo/0.jpg").convert('RGB')]
+video = pipe(prompt, input_images=images, num_inference_steps=50, seed=0, i2v_resolution=i2v_resolution)
+save_video(video, f"video_{i2v_resolution}.mp4", fps=30, quality=6)