<feat> add gradio app header.

app.py

@@ -0,0 +1,39 @@
+import os
+import sys
+import torch
+import diffusers
+import transformers
+import argparse
+import peft
+import copy
+import cv2
+import gradio as gr
+import numpy as np
+
+from peft import LoraConfig
+from omegaconf import OmegaConf
+from safetensors.torch import safe_open
+from PIL import Image, ImageDraw, ImageFilter
+
+from models import HunyuanVideoTransformer3DModel
+from pipelines import HunyuanVideoImageToVideoPipeline
+
+
+header = """
+# DRA-Ctrl Gradio App
+
+<div style="text-align: center; display: flex; justify-content: left; gap: 5px;">
+<a href="https://arxiv.org/pdf/2505.23325"><img src="https://img.shields.io/badge/ariXv-Paper-A42C25.svg" alt="arXiv"></a>
+<a href="https://huggingface.co/Kunbyte/DRA-Ctrl"><img src="https://img.shields.io/badge/🤗-Model-ffbd45.svg" alt="HuggingFace"></a>
+<a href="https://github.com/Kunbyte-AI/DRA-Ctrl"><img src="https://img.shields.io/badge/GitHub-Code-blue.svg?logo=github&" alt="GitHub"></a>
+<a href="https://dra-ctrl-2025.github.io/DRA-Ctrl/"><img src="https://img.shields.io/badge/Project-Page-blue" alt="Project"></a>
+</div>
+"""
+
+def create_app():
+    with gr.Blocks() as app:
+        gr.Markdown(header, elem_id="header")
+
+
+if __name__ == "__main__":
+    create_app().launch(debug=True, ssr_mode=False)

models/__init__.py

@@ -0,0 +1 @@
+from .hyvideo.transformer_hunyuan_video_i2v import HunyuanVideoTransformer3DModel

models/hyvideo/transformer_hunyuan_video_i2v.py

@@ -0,0 +1,1238 @@
+# Copyright 2024 The Hunyuan Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Modified by [Hengyuan Cao] in 2025.
+
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.loaders import FromOriginalModelMixin
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import PeftAdapterMixin
+from diffusers.utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.attention import FeedForward
+from diffusers.models.attention_processor import Attention, AttentionProcessor
+from diffusers.models.cache_utils import CacheMixin
+from diffusers.models.embeddings import (
+    CombinedTimestepTextProjEmbeddings,
+    PixArtAlphaTextProjection,
+    TimestepEmbedding,
+    Timesteps,
+    get_1d_rotary_pos_embed,
+)
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle, FP32LayerNorm
+from torch.nn.utils.rnn import pad_sequence
+
+try:
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    FLASH_ATTN_AVALIABLE = True
+except:
+    FLASH_ATTN_AVALIABLE = False
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class HunyuanVideoAttnProcessor2_0:
+    def __init__(self, inference_subject_driven: bool = False):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError(
+                "HunyuanVideoAttnProcessor2_0 requires PyTorch 2.0. To use it, please upgrade PyTorch to 2.0."
+            )
+        self.inference_subject_driven = inference_subject_driven
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if attn.add_q_proj is None and encoder_hidden_states is not None:
+            hidden_states = torch.cat([hidden_states, encoder_hidden_states], dim=1)
+
+        # 1. QKV projections
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        query = query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        key = key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+        value = value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+        # 2. QK normalization
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # 3. Rotational positional embeddings applied to latent stream
+        if image_rotary_emb is not None:
+            from diffusers.models.embeddings import apply_rotary_emb
+
+            if attn.add_q_proj is None and encoder_hidden_states is not None:
+                query = torch.cat(
+                    [
+                        apply_rotary_emb(query[:, :, : -encoder_hidden_states.shape[1]], image_rotary_emb),
+                        query[:, :, -encoder_hidden_states.shape[1] :],
+                    ],
+                    dim=2,
+                )
+                key = torch.cat(
+                    [
+                        apply_rotary_emb(key[:, :, : -encoder_hidden_states.shape[1]], image_rotary_emb),
+                        key[:, :, -encoder_hidden_states.shape[1] :],
+                    ],
+                    dim=2,
+                )
+            else:
+                query = apply_rotary_emb(query, image_rotary_emb)
+                key = apply_rotary_emb(key, image_rotary_emb)
+
+        # 4. Encoder condition QKV projection and normalization
+        if attn.add_q_proj is not None and encoder_hidden_states is not None:
+            encoder_query = attn.add_q_proj(encoder_hidden_states)
+            encoder_key = attn.add_k_proj(encoder_hidden_states)
+            encoder_value = attn.add_v_proj(encoder_hidden_states)
+
+            encoder_query = encoder_query.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+            encoder_key = encoder_key.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+            encoder_value = encoder_value.unflatten(2, (attn.heads, -1)).transpose(1, 2)
+
+            if attn.norm_added_q is not None:
+                encoder_query = attn.norm_added_q(encoder_query)
+            if attn.norm_added_k is not None:
+                encoder_key = attn.norm_added_k(encoder_key)
+
+            query = torch.cat([query, encoder_query], dim=2)
+            key = torch.cat([key, encoder_key], dim=2)
+            value = torch.cat([value, encoder_value], dim=2)
+
+        query = query.transpose(1, 2)   # batch, sequence, num_head, head_dim
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+
+        # 5. Attention
+        if FLASH_ATTN_AVALIABLE:
+            if attention_mask is None:
+                hidden_states = flash_attn_func(query, key, value, dropout=0.0)
+            else:
+                B, S, H, D = query.size()
+                unit_img_seq_len = 1024
+                unit_txt_seq_len = 144 + 252
+                if not (unit_img_seq_len*4+unit_txt_seq_len == S or 
+                    unit_img_seq_len*4+unit_txt_seq_len*2 == S):
+                    raise ValueError("Get wrong sequence length.")
+                if S == unit_img_seq_len*4+unit_txt_seq_len:
+                    seg_start = [0, unit_img_seq_len, unit_img_seq_len*4]
+                    seg_end = [unit_img_seq_len, unit_img_seq_len*4, unit_img_seq_len*4+unit_txt_seq_len]
+                    k_segs = [[0], [0, 1, 2], [1, 2]]
+                elif S == unit_img_seq_len*4+unit_txt_seq_len*2:
+                    seg_start = [0, unit_img_seq_len, unit_img_seq_len*4, unit_img_seq_len*4+unit_txt_seq_len]
+                    seg_end = [unit_img_seq_len, unit_img_seq_len*4, unit_img_seq_len*4+unit_txt_seq_len, S]
+                    k_segs = [[0, 3], [0, 1, 2], [1,2], [0, 3]]
+                valid_indices = attention_mask[:, 0, 0]
+                q_lens = torch.tensor([u[i:j].long().sum().item() for u in valid_indices for i,j in zip(seg_start, seg_end)], 
+                                        dtype=torch.int32, device=valid_indices.device)
+                k_lens = torch.tensor([sum([u[seg_start[seg]:seg_end[seg]].long().sum().item() for seg in segs]) for u in valid_indices for segs in k_segs], 
+                                        dtype=torch.int32, device=valid_indices.device)
+                query = torch.cat([u[i:j][v[i:j]] for u,v in zip(query, valid_indices) for i,j in zip(seg_start, seg_end)], dim=0)
+                if self.inference_subject_driven:
+                    key = torch.cat([torch.cat([ torch.cat([u[seg_start[seg]:seg_end[seg]][v[seg_start[seg]:seg_end[seg]]][:144], u[seg_start[seg]:seg_end[seg]][v[seg_start[seg]:seg_end[seg]]][144:] + 0.6 * u[seg_start[seg]:seg_end[seg]][v[seg_start[seg]:seg_end[seg]]][144:].abs().mean()], dim=0) if segs == [0, 1, 2] and seg == 2 else u[seg_start[seg]:seg_end[seg]][v[seg_start[seg]:seg_end[seg]]] for seg in segs], dim=0) \
+                                        for u,v in zip(key, valid_indices) for segs in k_segs], dim=0)
+                else:
+                    key = torch.cat([torch.cat([u[seg_start[seg]:seg_end[seg]][v[seg_start[seg]:seg_end[seg]]] for seg in segs], dim=0) \
+                                    for u,v in zip(key, valid_indices) for segs in k_segs], dim=0)
+                value = torch.cat([torch.cat([u[seg_start[seg]:seg_end[seg]][v[seg_start[seg]:seg_end[seg]]] for seg in segs], dim=0) \
+                                    for u,v in zip(value, valid_indices) for segs in k_segs], dim=0)
+                cu_seqlens_q = F.pad(q_lens.cumsum(dim=0), (1, 0)).to(torch.int32)
+                cu_seqlens_k = F.pad(k_lens.cumsum(dim=0), (1, 0)).to(torch.int32)
+                max_seqlen_q = torch.max(q_lens).item()
+                max_seqlen_k = torch.max(k_lens).item()
+                hidden_states = flash_attn_varlen_func(query, key, value, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k)
+                num_seq_parts = len(k_segs)
+                hidden_states = pad_sequence([
+                    hidden_states[start: end]
+                    for start, end in zip(cu_seqlens_q[::num_seq_parts][:-1], cu_seqlens_q[::num_seq_parts][1:])
+                ], batch_first=True)
+                hidden_states = F.pad(
+                    hidden_states,
+                    (0, 0, 0, 0, 0, S - hidden_states.size(1), 0, 0)
+                )
+        else:
+            query = query.permute(0, 2, 1, 3)   # batch, num_head, sequence, head_dim
+            key = key.permute(0, 2, 1, 3)
+            value = value.permute(0, 2, 1, 3)
+            hidden_states = F.scaled_dot_product_attention(
+                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            ) # use sdpa in torch may generate black output, upgrade to >=2.5.1 may solve this
+            hidden_states = hidden_states.transpose(1, 2)
+
+        # flatten num_head * head_dim
+        hidden_states = hidden_states.flatten(2, 3)
+        hidden_states = hidden_states.to(query.dtype)
+
+
+        # 6. Output projection
+        if encoder_hidden_states is not None:
+            hidden_states, encoder_hidden_states = (
+                hidden_states[:, : -encoder_hidden_states.shape[1]],
+                hidden_states[:, -encoder_hidden_states.shape[1] :],
+            )
+
+            if getattr(attn, "to_out", None) is not None:
+                hidden_states = attn.to_out[0](hidden_states)
+                hidden_states = attn.to_out[1](hidden_states)
+
+            if getattr(attn, "to_add_out", None) is not None:
+                encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        return hidden_states, encoder_hidden_states
+
+
+class HunyuanVideoPatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size: Union[int, Tuple[int, int, int]] = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ) -> None:
+        super().__init__()
+
+        patch_size = (patch_size, patch_size, patch_size) if isinstance(patch_size, int) else patch_size
+        self.proj = nn.Conv3d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.proj(hidden_states)
+        hidden_states = hidden_states.flatten(2).transpose(1, 2)  # BCFHW -> BNC
+        return hidden_states
+
+
+class HunyuanVideoAdaNorm(nn.Module):
+    def __init__(self, in_features: int, out_features: Optional[int] = None) -> None:
+        super().__init__()
+
+        out_features = out_features or 2 * in_features
+        self.linear = nn.Linear(in_features, out_features)
+        self.nonlinearity = nn.SiLU()
+
+    def forward(
+        self, temb: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        temb = self.linear(self.nonlinearity(temb))
+        gate_msa, gate_mlp = temb.chunk(2, dim=1)
+        gate_msa, gate_mlp = gate_msa.unsqueeze(1), gate_mlp.unsqueeze(1)
+        return gate_msa, gate_mlp
+
+
+class HunyuanVideoTokenReplaceAdaLayerNormZero(nn.Module):
+    def __init__(self, embedding_dim: int, norm_type: str = "layer_norm", bias: bool = True):
+        super().__init__()
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=bias)
+
+        if norm_type == "layer_norm":
+            self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
+        elif norm_type == "fp32_layer_norm":
+            self.norm = FP32LayerNorm(embedding_dim, elementwise_affine=False, bias=False)
+        else:
+            raise ValueError(
+                f"Unsupported `norm_type` ({norm_type}) provided. Supported ones are: 'layer_norm', 'fp32_layer_norm'."
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        emb: torch.Tensor,
+        token_replace_emb: torch.Tensor,
+        first_frame_num_tokens: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        emb = self.linear(self.silu(emb))
+        token_replace_emb = self.linear(self.silu(token_replace_emb))
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1)
+        tr_shift_msa, tr_scale_msa, tr_gate_msa, tr_shift_mlp, tr_scale_mlp, tr_gate_mlp = token_replace_emb.chunk(
+            6, dim=1
+        )
+
+        norm_hidden_states = self.norm(hidden_states)
+        hidden_states_zero = (
+            norm_hidden_states[:, :first_frame_num_tokens] * (1 + tr_scale_msa[:, None]) + tr_shift_msa[:, None]
+        )
+        hidden_states_orig = (
+            norm_hidden_states[:, first_frame_num_tokens:] * (1 + scale_msa[:, None]) + shift_msa[:, None]
+        )
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+
+        return (
+            hidden_states,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            tr_gate_msa,
+            tr_shift_mlp,
+            tr_scale_mlp,
+            tr_gate_mlp,
+        )
+
+
+class HunyuanVideoTokenReplaceAdaLayerNormZeroSingle(nn.Module):
+    def __init__(self, embedding_dim: int, norm_type: str = "layer_norm", bias: bool = True):
+        super().__init__()
+
+        self.silu = nn.SiLU()
+        self.linear = nn.Linear(embedding_dim, 3 * embedding_dim, bias=bias)
+
+        if norm_type == "layer_norm":
+            self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
+        else:
+            raise ValueError(
+                f"Unsupported `norm_type` ({norm_type}) provided. Supported ones are: 'layer_norm', 'fp32_layer_norm'."
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        emb: torch.Tensor,
+        token_replace_emb: torch.Tensor,
+        first_frame_num_tokens: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        emb = self.linear(self.silu(emb))
+        token_replace_emb = self.linear(self.silu(token_replace_emb))
+
+        shift_msa, scale_msa, gate_msa = emb.chunk(3, dim=1)
+        tr_shift_msa, tr_scale_msa, tr_gate_msa = token_replace_emb.chunk(3, dim=1)
+
+        norm_hidden_states = self.norm(hidden_states)
+        hidden_states_zero = (
+            norm_hidden_states[:, :first_frame_num_tokens] * (1 + tr_scale_msa[:, None]) + tr_shift_msa[:, None]
+        )
+        hidden_states_orig = (
+            norm_hidden_states[:, first_frame_num_tokens:] * (1 + scale_msa[:, None]) + shift_msa[:, None]
+        )
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+
+        return hidden_states, gate_msa, tr_gate_msa
+
+
+class HunyuanVideoConditionEmbedding(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        pooled_projection_dim: int,
+        guidance_embeds: bool,
+        image_condition_type: Optional[str] = None,
+    ):
+        super().__init__()
+
+        self.image_condition_type = image_condition_type
+
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+        self.text_embedder = PixArtAlphaTextProjection(pooled_projection_dim, embedding_dim, act_fn="silu")
+
+        self.guidance_embedder = None
+        if guidance_embeds:
+            self.guidance_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+    def forward(
+        self, timestep: torch.Tensor, pooled_projection: torch.Tensor, guidance: Optional[torch.Tensor] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=pooled_projection.dtype))  # (N, D)
+        pooled_projections = self.text_embedder(pooled_projection)
+        conditioning = timesteps_emb + pooled_projections
+
+        token_replace_emb = None
+        if self.image_condition_type == "token_replace":
+            token_replace_timestep = torch.zeros_like(timestep)
+            token_replace_proj = self.time_proj(token_replace_timestep)
+            token_replace_emb = self.timestep_embedder(token_replace_proj.to(dtype=pooled_projection.dtype))
+            token_replace_emb = token_replace_emb + pooled_projections
+
+        if self.guidance_embedder is not None:
+            guidance_proj = self.time_proj(guidance)
+            guidance_emb = self.guidance_embedder(guidance_proj.to(dtype=pooled_projection.dtype))
+            conditioning = conditioning + guidance_emb
+
+        return conditioning, token_replace_emb
+
+
+class HunyuanVideoIndividualTokenRefinerBlock(nn.Module):
+    def __init__(
+        self,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        mlp_width_ratio: str = 4.0,
+        mlp_drop_rate: float = 0.0,
+        attention_bias: bool = True,
+    ) -> None:
+        super().__init__()
+
+        hidden_size = num_attention_heads * attention_head_dim
+
+        self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6)
+        self.attn = Attention(
+            query_dim=hidden_size,
+            cross_attention_dim=None,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            bias=attention_bias,
+        )
+
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=True, eps=1e-6)
+        self.ff = FeedForward(hidden_size, mult=mlp_width_ratio, activation_fn="linear-silu", dropout=mlp_drop_rate)
+
+        self.norm_out = HunyuanVideoAdaNorm(hidden_size, 2 * hidden_size)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        norm_hidden_states = self.norm1(hidden_states)
+
+        attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=None,
+            attention_mask=attention_mask,
+        )
+
+        gate_msa, gate_mlp = self.norm_out(temb)
+        hidden_states = hidden_states + attn_output * gate_msa
+
+        ff_output = self.ff(self.norm2(hidden_states))
+        hidden_states = hidden_states + ff_output * gate_mlp
+
+        return hidden_states
+
+
+class HunyuanVideoIndividualTokenRefiner(nn.Module):
+    def __init__(
+        self,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_layers: int,
+        mlp_width_ratio: float = 4.0,
+        mlp_drop_rate: float = 0.0,
+        attention_bias: bool = True,
+    ) -> None:
+        super().__init__()
+
+        self.refiner_blocks = nn.ModuleList(
+            [
+                HunyuanVideoIndividualTokenRefinerBlock(
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    mlp_width_ratio=mlp_width_ratio,
+                    mlp_drop_rate=mlp_drop_rate,
+                    attention_bias=attention_bias,
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> None:
+        self_attn_mask = None
+        if attention_mask is not None:
+            batch_size = attention_mask.shape[0]
+            seq_len = attention_mask.shape[1]
+            attention_mask = attention_mask.to(hidden_states.device).bool()
+            self_attn_mask_1 = attention_mask.view(batch_size, 1, 1, seq_len).repeat(1, 1, seq_len, 1)
+            self_attn_mask_2 = self_attn_mask_1.transpose(2, 3)
+            self_attn_mask = (self_attn_mask_1 & self_attn_mask_2).bool()
+            self_attn_mask[:, :, :, 0] = True
+
+        for block in self.refiner_blocks:
+            hidden_states = block(hidden_states, temb, self_attn_mask)
+
+        return hidden_states
+
+
+class HunyuanVideoTokenRefiner(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_layers: int,
+        mlp_ratio: float = 4.0,
+        mlp_drop_rate: float = 0.0,
+        attention_bias: bool = True,
+    ) -> None:
+        super().__init__()
+
+        hidden_size = num_attention_heads * attention_head_dim
+
+        self.time_text_embed = CombinedTimestepTextProjEmbeddings(
+            embedding_dim=hidden_size, pooled_projection_dim=in_channels
+        )
+        self.proj_in = nn.Linear(in_channels, hidden_size, bias=True)
+        self.token_refiner = HunyuanVideoIndividualTokenRefiner(
+            num_attention_heads=num_attention_heads,
+            attention_head_dim=attention_head_dim,
+            num_layers=num_layers,
+            mlp_width_ratio=mlp_ratio,
+            mlp_drop_rate=mlp_drop_rate,
+            attention_bias=attention_bias,
+        )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+    ) -> torch.Tensor:
+        if attention_mask is None:
+            pooled_projections = hidden_states.mean(dim=1)
+        else:
+            original_dtype = hidden_states.dtype
+            mask_float = attention_mask.float().unsqueeze(-1)
+            pooled_projections = (hidden_states * mask_float).sum(dim=1) / mask_float.sum(dim=1)
+            pooled_projections = pooled_projections.to(original_dtype)
+
+        temb = self.time_text_embed(timestep, pooled_projections)
+        hidden_states = self.proj_in(hidden_states)
+        hidden_states = self.token_refiner(hidden_states, temb, attention_mask)
+
+        return hidden_states
+
+
+class HunyuanVideoRotaryPosEmbed(nn.Module):
+    def __init__(self, patch_size: int, patch_size_t: int, rope_dim: List[int], theta: float = 256.0) -> None:
+        super().__init__()
+
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.rope_dim = rope_dim
+        self.theta = theta
+
+    def forward(self, hidden_states: torch.Tensor, frame_gap: Union[int, None] = None) -> torch.Tensor:
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        rope_sizes = [num_frames // self.patch_size_t, height // self.patch_size, width // self.patch_size]
+
+        axes_grids = []
+        for i in range(3):
+            # Note: The following line diverges from original behaviour. We create the grid on the device, whereas
+            # original implementation creates it on CPU and then moves it to device. This results in numerical
+            # differences in layerwise debugging outputs, but visually it is the same.
+            grid = torch.arange(0, rope_sizes[i], device=hidden_states.device, dtype=torch.float32)
+            if frame_gap is not None and i == 0:
+                grid = grid * frame_gap
+            axes_grids.append(grid)
+        grid = torch.meshgrid(*axes_grids, indexing="ij")  # [W, H, T]
+        grid = torch.stack(grid, dim=0)  # [3, W, H, T]
+
+        freqs = []
+        for i in range(3):
+            freq = get_1d_rotary_pos_embed(self.rope_dim[i], grid[i].reshape(-1), self.theta, use_real=True)
+            freqs.append(freq)
+
+        freqs_cos = torch.cat([f[0] for f in freqs], dim=1)  # (W * H * T, D / 2)
+        freqs_sin = torch.cat([f[1] for f in freqs], dim=1)  # (W * H * T, D / 2)
+        return freqs_cos, freqs_sin
+
+
+class HunyuanVideoSingleTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        mlp_ratio: float = 4.0,
+        qk_norm: str = "rms_norm",
+        inference_subject_driven: bool = False,
+    ) -> None:
+        super().__init__()
+
+        hidden_size = num_attention_heads * attention_head_dim
+        mlp_dim = int(hidden_size * mlp_ratio)
+
+        self.attn = Attention(
+            query_dim=hidden_size,
+            cross_attention_dim=None,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=hidden_size,
+            bias=True,
+            processor=HunyuanVideoAttnProcessor2_0(inference_subject_driven=inference_subject_driven),
+            qk_norm=qk_norm,
+            eps=1e-6,
+            pre_only=True,
+        )
+
+        self.norm = AdaLayerNormZeroSingle(hidden_size, norm_type="layer_norm")
+        self.proj_mlp = nn.Linear(hidden_size, mlp_dim)
+        self.act_mlp = nn.GELU(approximate="tanh")
+        self.proj_out = nn.Linear(hidden_size + mlp_dim, hidden_size)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        text_seq_length = encoder_hidden_states.shape[1]
+        hidden_states = torch.cat([hidden_states, encoder_hidden_states], dim=1)
+
+        residual = hidden_states
+
+        # 1. Input normalization
+        norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
+        mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
+
+        norm_hidden_states, norm_encoder_hidden_states = (
+            norm_hidden_states[:, :-text_seq_length, :],
+            norm_hidden_states[:, -text_seq_length:, :],
+        )
+
+        # 2. Attention
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            attention_mask=attention_mask,
+            image_rotary_emb=image_rotary_emb,
+        )
+        attn_output = torch.cat([attn_output, context_attn_output], dim=1)
+
+        # 3. Modulation and residual connection
+        hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
+        hidden_states = gate.unsqueeze(1) * self.proj_out(hidden_states)
+        hidden_states = hidden_states + residual
+
+        hidden_states, encoder_hidden_states = (
+            hidden_states[:, :-text_seq_length, :],
+            hidden_states[:, -text_seq_length:, :],
+        )
+        return hidden_states, encoder_hidden_states
+
+
+class HunyuanVideoTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        mlp_ratio: float,
+        qk_norm: str = "rms_norm",
+        inference_subject_driven: bool = False,
+    ) -> None:
+        super().__init__()
+
+        hidden_size = num_attention_heads * attention_head_dim
+
+        self.norm1 = AdaLayerNormZero(hidden_size, norm_type="layer_norm")
+        self.norm1_context = AdaLayerNormZero(hidden_size, norm_type="layer_norm")
+
+        self.attn = Attention(
+            query_dim=hidden_size,
+            cross_attention_dim=None,
+            added_kv_proj_dim=hidden_size,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=hidden_size,
+            context_pre_only=False,
+            bias=True,
+            processor=HunyuanVideoAttnProcessor2_0(inference_subject_driven=inference_subject_driven),
+            qk_norm=qk_norm,
+            eps=1e-6,
+        )
+
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.ff = FeedForward(hidden_size, mult=mlp_ratio, activation_fn="gelu-approximate")
+
+        self.norm2_context = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.ff_context = FeedForward(hidden_size, mult=mlp_ratio, activation_fn="gelu-approximate")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        *args,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # 1. Input normalization
+        norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
+        norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+            encoder_hidden_states, emb=temb
+        )
+
+        # 2. Joint attention
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            attention_mask=attention_mask,
+            image_rotary_emb=freqs_cis,
+        )
+
+        # 3. Modulation and residual connection
+        hidden_states = hidden_states + attn_output * gate_msa.unsqueeze(1)
+        encoder_hidden_states = encoder_hidden_states + context_attn_output * c_gate_msa.unsqueeze(1)
+
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+        norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+
+        # 4. Feed-forward
+        ff_output = self.ff(norm_hidden_states)
+        context_ff_output = self.ff_context(norm_encoder_hidden_states)
+
+        hidden_states = hidden_states + gate_mlp.unsqueeze(1) * ff_output
+        encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+
+        return hidden_states, encoder_hidden_states
+
+
+class HunyuanVideoTokenReplaceSingleTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        mlp_ratio: float = 4.0,
+        qk_norm: str = "rms_norm",
+        inference_subject_driven: bool = False,
+    ) -> None:
+        super().__init__()
+
+        hidden_size = num_attention_heads * attention_head_dim
+        mlp_dim = int(hidden_size * mlp_ratio)
+
+        self.attn = Attention(
+            query_dim=hidden_size,
+            cross_attention_dim=None,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=hidden_size,
+            bias=True,
+            processor=HunyuanVideoAttnProcessor2_0(inference_subject_driven=inference_subject_driven),
+            qk_norm=qk_norm,
+            eps=1e-6,
+            pre_only=True,
+        )
+
+        self.norm = HunyuanVideoTokenReplaceAdaLayerNormZeroSingle(hidden_size, norm_type="layer_norm")
+        self.proj_mlp = nn.Linear(hidden_size, mlp_dim)
+        self.act_mlp = nn.GELU(approximate="tanh")
+        self.proj_out = nn.Linear(hidden_size + mlp_dim, hidden_size)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        token_replace_emb: torch.Tensor = None,
+        num_tokens: int = None,
+    ) -> torch.Tensor:
+        text_seq_length = encoder_hidden_states.shape[1]
+        hidden_states = torch.cat([hidden_states, encoder_hidden_states], dim=1)
+
+        residual = hidden_states
+
+        # 1. Input normalization
+        norm_hidden_states, gate, tr_gate = self.norm(hidden_states, temb, token_replace_emb, num_tokens)
+        mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
+
+        norm_hidden_states, norm_encoder_hidden_states = (
+            norm_hidden_states[:, :-text_seq_length, :],
+            norm_hidden_states[:, -text_seq_length:, :],
+        )
+
+        # 2. Attention
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            attention_mask=attention_mask,
+            image_rotary_emb=image_rotary_emb,
+        )
+        attn_output = torch.cat([attn_output, context_attn_output], dim=1)
+
+        # 3. Modulation and residual connection
+        hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
+
+        proj_output = self.proj_out(hidden_states)
+        hidden_states_zero = proj_output[:, :num_tokens] * tr_gate.unsqueeze(1)
+        hidden_states_orig = proj_output[:, num_tokens:] * gate.unsqueeze(1)
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+        hidden_states = hidden_states + residual
+
+        hidden_states, encoder_hidden_states = (
+            hidden_states[:, :-text_seq_length, :],
+            hidden_states[:, -text_seq_length:, :],
+        )
+        return hidden_states, encoder_hidden_states
+
+
+class HunyuanVideoTokenReplaceTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        mlp_ratio: float,
+        qk_norm: str = "rms_norm",
+        inference_subject_driven: bool = False,
+    ) -> None:
+        super().__init__()
+
+        hidden_size = num_attention_heads * attention_head_dim
+
+        self.norm1 = HunyuanVideoTokenReplaceAdaLayerNormZero(hidden_size, norm_type="layer_norm")
+        self.norm1_context = AdaLayerNormZero(hidden_size, norm_type="layer_norm")
+
+        self.attn = Attention(
+            query_dim=hidden_size,
+            cross_attention_dim=None,
+            added_kv_proj_dim=hidden_size,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=hidden_size,
+            context_pre_only=False,
+            bias=True,
+            processor=HunyuanVideoAttnProcessor2_0(inference_subject_driven=inference_subject_driven),
+            qk_norm=qk_norm,
+            eps=1e-6,
+        )
+
+        self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.ff = FeedForward(hidden_size, mult=mlp_ratio, activation_fn="gelu-approximate")
+
+        self.norm2_context = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.ff_context = FeedForward(hidden_size, mult=mlp_ratio, activation_fn="gelu-approximate")
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        temb: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        freqs_cis: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        token_replace_emb: torch.Tensor = None,
+        num_tokens: int = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # 1. Input normalization
+        (
+            norm_hidden_states,
+            gate_msa,
+            shift_mlp,
+            scale_mlp,
+            gate_mlp,
+            tr_gate_msa,
+            tr_shift_mlp,
+            tr_scale_mlp,
+            tr_gate_mlp,
+        ) = self.norm1(hidden_states, temb, token_replace_emb, num_tokens)
+        norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+            encoder_hidden_states, emb=temb
+        )
+
+        # 2. Joint attention
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            attention_mask=attention_mask,
+            image_rotary_emb=freqs_cis,
+        )
+
+        # 3. Modulation and residual connection
+        hidden_states_zero = hidden_states[:, :num_tokens] + attn_output[:, :num_tokens] * tr_gate_msa.unsqueeze(1)
+        hidden_states_orig = hidden_states[:, num_tokens:] + attn_output[:, num_tokens:] * gate_msa.unsqueeze(1)
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+        encoder_hidden_states = encoder_hidden_states + context_attn_output * c_gate_msa.unsqueeze(1)
+
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+
+        hidden_states_zero = norm_hidden_states[:, :num_tokens] * (1 + tr_scale_mlp[:, None]) + tr_shift_mlp[:, None]
+        hidden_states_orig = norm_hidden_states[:, num_tokens:] * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+        norm_hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+        norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+
+        # 4. Feed-forward
+        ff_output = self.ff(norm_hidden_states)
+        context_ff_output = self.ff_context(norm_encoder_hidden_states)
+
+        hidden_states_zero = hidden_states[:, :num_tokens] + ff_output[:, :num_tokens] * tr_gate_mlp.unsqueeze(1)
+        hidden_states_orig = hidden_states[:, num_tokens:] + ff_output[:, num_tokens:] * gate_mlp.unsqueeze(1)
+        hidden_states = torch.cat([hidden_states_zero, hidden_states_orig], dim=1)
+        encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+
+        return hidden_states, encoder_hidden_states
+
+
+class HunyuanVideoTransformer3DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin):
+    r"""
+    A Transformer model for video-like data used in [HunyuanVideo](https://huggingface.co/tencent/HunyuanVideo).
+
+    Args:
+        in_channels (`int`, defaults to `16`):
+            The number of channels in the input.
+        out_channels (`int`, defaults to `16`):
+            The number of channels in the output.
+        num_attention_heads (`int`, defaults to `24`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, defaults to `128`):
+            The number of channels in each head.
+        num_layers (`int`, defaults to `20`):
+            The number of layers of dual-stream blocks to use.
+        num_single_layers (`int`, defaults to `40`):
+            The number of layers of single-stream blocks to use.
+        num_refiner_layers (`int`, defaults to `2`):
+            The number of layers of refiner blocks to use.
+        mlp_ratio (`float`, defaults to `4.0`):
+            The ratio of the hidden layer size to the input size in the feedforward network.
+        patch_size (`int`, defaults to `2`):
+            The size of the spatial patches to use in the patch embedding layer.
+        patch_size_t (`int`, defaults to `1`):
+            The size of the tmeporal patches to use in the patch embedding layer.
+        qk_norm (`str`, defaults to `rms_norm`):
+            The normalization to use for the query and key projections in the attention layers.
+        guidance_embeds (`bool`, defaults to `True`):
+            Whether to use guidance embeddings in the model.
+        text_embed_dim (`int`, defaults to `4096`):
+            Input dimension of text embeddings from the text encoder.
+        pooled_projection_dim (`int`, defaults to `768`):
+            The dimension of the pooled projection of the text embeddings.
+        rope_theta (`float`, defaults to `256.0`):
+            The value of theta to use in the RoPE layer.
+        rope_axes_dim (`Tuple[int]`, defaults to `(16, 56, 56)`):
+            The dimensions of the axes to use in the RoPE layer.
+        image_condition_type (`str`, *optional*, defaults to `None`):
+            The type of image conditioning to use. If `None`, no image conditioning is used. If `latent_concat`, the
+            image is concatenated to the latent stream. If `token_replace`, the image is used to replace first-frame
+            tokens in the latent stream and apply conditioning.
+    """
+
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["x_embedder", "context_embedder", "norm"]
+    _no_split_modules = [
+        "HunyuanVideoTransformerBlock",
+        "HunyuanVideoSingleTransformerBlock",
+        "HunyuanVideoPatchEmbed",
+        "HunyuanVideoTokenRefiner",
+    ]
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        num_attention_heads: int = 24,
+        attention_head_dim: int = 128,
+        num_layers: int = 20,
+        num_single_layers: int = 40,
+        num_refiner_layers: int = 2,
+        mlp_ratio: float = 4.0,
+        patch_size: int = 2,
+        patch_size_t: int = 1,
+        qk_norm: str = "rms_norm",
+        guidance_embeds: bool = True,
+        text_embed_dim: int = 4096,
+        pooled_projection_dim: int = 768,
+        rope_theta: float = 256.0,
+        rope_axes_dim: Tuple[int] = (16, 56, 56),
+        image_condition_type: Optional[str] = None,
+        inference_subject_driven: bool = False,
+    ) -> None:
+        super().__init__()
+
+        supported_image_condition_types = ["latent_concat", "token_replace"]
+        if image_condition_type is not None and image_condition_type not in supported_image_condition_types:
+            raise ValueError(
+                f"Invalid `image_condition_type` ({image_condition_type}). Supported ones are: {supported_image_condition_types}"
+            )
+
+        inner_dim = num_attention_heads * attention_head_dim
+        out_channels = out_channels or in_channels
+
+        # 1. Latent and condition embedders
+        self.x_embedder = HunyuanVideoPatchEmbed((patch_size_t, patch_size, patch_size), in_channels, inner_dim)
+        self.context_embedder = HunyuanVideoTokenRefiner(
+            text_embed_dim, num_attention_heads, attention_head_dim, num_layers=num_refiner_layers
+        )
+
+        self.time_text_embed = HunyuanVideoConditionEmbedding(
+            inner_dim, pooled_projection_dim, guidance_embeds, image_condition_type
+        )
+
+        # 2. RoPE
+        self.rope = HunyuanVideoRotaryPosEmbed(patch_size, patch_size_t, rope_axes_dim, rope_theta)
+
+        # 3. Dual stream transformer blocks
+        if image_condition_type == "token_replace":
+            self.transformer_blocks = nn.ModuleList(
+                [
+                    HunyuanVideoTokenReplaceTransformerBlock(
+                        num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm, inference_subject_driven=inference_subject_driven
+                    )
+                    for _ in range(num_layers)
+                ]
+            )
+        else:
+            self.transformer_blocks = nn.ModuleList(
+                [
+                    HunyuanVideoTransformerBlock(
+                        num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm, inference_subject_driven=inference_subject_driven
+                    )
+                    for _ in range(num_layers)
+                ]
+            )
+
+        # 4. Single stream transformer blocks
+        if image_condition_type == "token_replace":
+            self.single_transformer_blocks = nn.ModuleList(
+                [
+                    HunyuanVideoTokenReplaceSingleTransformerBlock(
+                        num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm, inference_subject_driven=inference_subject_driven
+                    )
+                    for _ in range(num_single_layers)
+                ]
+            )
+        else:
+            self.single_transformer_blocks = nn.ModuleList(
+                [
+                    HunyuanVideoSingleTransformerBlock(
+                        num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm, inference_subject_driven=inference_subject_driven
+                    )
+                    for _ in range(num_single_layers)
+                ]
+            )
+
+        # 5. Output projection
+        self.norm_out = AdaLayerNormContinuous(inner_dim, inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(inner_dim, patch_size_t * patch_size * patch_size * out_channels)
+
+        self.gradient_checkpointing = False
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_attention_mask: torch.Tensor,
+        pooled_projections: torch.Tensor,
+        encoder_hidden_states_condition: Union[torch.Tensor, None] = None,
+        encoder_attention_mask_condition: Union[torch.Tensor, None] = None,
+        guidance: torch.Tensor = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+        frame_gap: Union[int, None] = None,
+    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
+        batch_size, num_channels, num_frames, height, width = hidden_states.shape
+        p, p_t = self.config.patch_size, self.config.patch_size_t
+        post_patch_num_frames = num_frames // p_t
+        post_patch_height = height // p
+        post_patch_width = width // p
+        first_frame_num_tokens = 1 * post_patch_height * post_patch_width
+
+        # 1. RoPE
+        image_rotary_emb = self.rope(hidden_states, frame_gap=frame_gap)
+
+        # 2. Conditional embeddings
+        temb, token_replace_emb = self.time_text_embed(timestep, pooled_projections, guidance)
+
+        hidden_states = self.x_embedder(hidden_states)
+        encoder_hidden_states = self.context_embedder(encoder_hidden_states, timestep, encoder_attention_mask)
+        if encoder_hidden_states_condition is not None and encoder_attention_mask_condition is not None:
+            encoder_hidden_states_condition = self.context_embedder(
+                encoder_hidden_states_condition,
+                torch.zeros_like(timestep),
+                encoder_attention_mask_condition,
+            )
+            encoder_hidden_states = torch.cat([encoder_hidden_states, encoder_hidden_states_condition], dim=1)
+            encoder_attention_mask = torch.cat([encoder_attention_mask, encoder_attention_mask_condition], dim=1)
+
+        # 3. Attention mask preparation
+        latent_sequence_length = hidden_states.shape[1]
+        condition_sequence_length = encoder_hidden_states.shape[1]
+        sequence_length = latent_sequence_length + condition_sequence_length
+        attention_mask = torch.zeros(
+            batch_size, sequence_length, device=hidden_states.device, dtype=torch.bool
+        )  # [B, N]
+
+        effective_condition_sequence_length = encoder_attention_mask.sum(dim=1, dtype=torch.int)  # [B,]
+        effective_sequence_length = latent_sequence_length + effective_condition_sequence_length
+
+        for i in range(batch_size):
+            if encoder_attention_mask_condition is not None and encoder_attention_mask_condition is not None:
+                attention_mask[i, : latent_sequence_length] = True
+                attention_mask[i, latent_sequence_length :][encoder_attention_mask[i] == 1.] = True
+            else:
+                attention_mask[i, : effective_sequence_length[i]] = True
+        # [B, 1, 1, N], for broadcasting across attention heads
+        attention_mask = attention_mask.unsqueeze(1).unsqueeze(1)
+
+        # 4. Transformer blocks
+        if torch.is_grad_enabled() and self.gradient_checkpointing:
+            for block in self.transformer_blocks:
+                hidden_states, encoder_hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    attention_mask,
+                    image_rotary_emb,
+                    token_replace_emb,
+                    first_frame_num_tokens,
+                )
+
+            for block in self.single_transformer_blocks:
+                hidden_states, encoder_hidden_states = self._gradient_checkpointing_func(
+                    block,
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    attention_mask,
+                    image_rotary_emb,
+                    token_replace_emb,
+                    first_frame_num_tokens,
+                )
+
+        else:
+            for block in self.transformer_blocks:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    attention_mask,
+                    image_rotary_emb,
+                    token_replace_emb,
+                    first_frame_num_tokens,
+                )
+
+            for block in self.single_transformer_blocks:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states,
+                    encoder_hidden_states,
+                    temb,
+                    attention_mask,
+                    image_rotary_emb,
+                    token_replace_emb,
+                    first_frame_num_tokens,
+                )
+
+        # 5. Output projection
+        hidden_states = self.norm_out(hidden_states, temb)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states.reshape(
+            batch_size, post_patch_num_frames, post_patch_height, post_patch_width, -1, p_t, p, p
+        )
+        hidden_states = hidden_states.permute(0, 4, 1, 5, 2, 6, 3, 7)
+        hidden_states = hidden_states.flatten(6, 7).flatten(4, 5).flatten(2, 3)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (hidden_states,)
+
+        return Transformer2DModelOutput(sample=hidden_states)

pipelines/__init__.py

@@ -0,0 +1 @@
+from .pipeline_hunyuan_video_i2v import HunyuanVideoImageToVideoPipeline

pipelines/pipeline_hunyuan_video_i2v.py

@@ -0,0 +1,969 @@
+# Copyright 2024 The HunyuanVideo Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Modified by [Hengyuan Cao] in 2025.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import (
+    CLIPImageProcessor,
+    CLIPTextModel,
+    CLIPTokenizer,
+    LlamaTokenizerFast,
+    LlavaForConditionalGeneration,
+)
+
+from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
+from diffusers.loaders import HunyuanVideoLoraLoaderMixin
+from diffusers.models import AutoencoderKLHunyuanVideo, HunyuanVideoTransformer3DModel
+from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.video_processor import VideoProcessor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.hunyuan_video.pipeline_output import HunyuanVideoPipelineOutput
+
+
+if is_torch_xla_available():
+    import torch_xla.core.xla_model as xm
+
+    XLA_AVAILABLE = True
+else:
+    XLA_AVAILABLE = False
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```python
+        >>> import torch
+        >>> from diffusers import HunyuanVideoImageToVideoPipeline, HunyuanVideoTransformer3DModel
+        >>> from diffusers.utils import load_image, export_to_video
+
+        >>> # Available checkpoints: hunyuanvideo-community/HunyuanVideo-I2V, hunyuanvideo-community/HunyuanVideo-I2V-33ch
+        >>> model_id = "hunyuanvideo-community/HunyuanVideo-I2V"
+        >>> transformer = HunyuanVideoTransformer3DModel.from_pretrained(
+        ...     model_id, subfolder="transformer", torch_dtype=torch.bfloat16
+        ... )
+        >>> pipe = HunyuanVideoImageToVideoPipeline.from_pretrained(
+        ...     model_id, transformer=transformer, torch_dtype=torch.float16
+        ... )
+        >>> pipe.vae.enable_tiling()
+        >>> pipe.to("cuda")
+
+        >>> prompt = "A man with short gray hair plays a red electric guitar."
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png"
+        ... )
+
+        >>> # If using hunyuanvideo-community/HunyuanVideo-I2V
+        >>> output = pipe(image=image, prompt=prompt, guidance_scale=6.0).frames[0]
+
+        >>> # If using hunyuanvideo-community/HunyuanVideo-I2V-33ch
+        >>> output = pipe(image=image, prompt=prompt, guidance_scale=1.0, true_cfg_scale=1.0).frames[0]
+
+        >>> export_to_video(output, "output.mp4", fps=15)
+        ```
+"""
+
+
+DEFAULT_PROMPT_TEMPLATE = {
+    "template": (
+        "<|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"
+    ),
+    "crop_start": 103,
+    "image_emb_start": 5,
+    "image_emb_end": 581,
+    "image_emb_len": 576,
+    "double_return_token_id": 271,
+}
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    sigmas: Optional[List[float]] = None,
+    **kwargs,
+):
+    r"""
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
+            must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
+            `num_inference_steps` and `sigmas` must be `None`.
+        sigmas (`List[float]`, *optional*):
+            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
+            `num_inference_steps` and `timesteps` must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None and sigmas is not None:
+        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    elif sigmas is not None:
+        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accept_sigmas:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" sigmas schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+        return encoder_output.latent_dist.sample(generator)
+    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+        return encoder_output.latent_dist.mode()
+    elif hasattr(encoder_output, "latents"):
+        return encoder_output.latents
+    else:
+        raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class HunyuanVideoImageToVideoPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin):
+    r"""
+    Pipeline for image-to-video generation using HunyuanVideo.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    Args:
+        text_encoder ([`LlavaForConditionalGeneration`]):
+            [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers).
+        tokenizer (`LlamaTokenizer`):
+            Tokenizer from [Llava Llama3-8B](https://huggingface.co/xtuner/llava-llama-3-8b-v1_1-transformers).
+        transformer ([`HunyuanVideoTransformer3DModel`]):
+            Conditional Transformer to denoise the encoded image latents.
+        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
+            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+        vae ([`AutoencoderKLHunyuanVideo`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+        text_encoder_2 ([`CLIPTextModel`]):
+            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+        tokenizer_2 (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
+    """
+
+    model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        text_encoder: LlavaForConditionalGeneration,
+        tokenizer: LlamaTokenizerFast,
+        transformer: HunyuanVideoTransformer3DModel,
+        vae: AutoencoderKLHunyuanVideo,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        text_encoder_2: CLIPTextModel,
+        tokenizer_2: CLIPTokenizer,
+        image_processor: CLIPImageProcessor,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            transformer=transformer,
+            scheduler=scheduler,
+            text_encoder_2=text_encoder_2,
+            tokenizer_2=tokenizer_2,
+            image_processor=image_processor,
+        )
+
+        self.vae_scaling_factor = self.vae.config.scaling_factor if getattr(self, "vae", None) else 0.476986
+        self.vae_scale_factor_temporal = self.vae.temporal_compression_ratio if getattr(self, "vae", None) else 4
+        self.vae_scale_factor_spatial = self.vae.spatial_compression_ratio if getattr(self, "vae", None) else 8
+        self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
+
+    def _get_llama_prompt_embeds(
+        self,
+        image: torch.Tensor,
+        prompt: Union[str, List[str]],
+        prompt_template: Dict[str, Any],
+        num_videos_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 256,
+        num_hidden_layers_to_skip: int = 2,
+        image_embed_interleave: int = 2,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+        prompt = [prompt_template["template"].format(p) for p in prompt]
+
+        crop_start = prompt_template.get("crop_start", None)
+        if crop_start is None:
+            prompt_template_input = self.tokenizer(
+                prompt_template["template"],
+                padding="max_length",
+                return_tensors="pt",
+                return_length=False,
+                return_overflowing_tokens=False,
+                return_attention_mask=False,
+            )
+            crop_start = prompt_template_input["input_ids"].shape[-1]
+            # Remove <|start_header_id|>, <|end_header_id|>, assistant, <|eot_id|>, and placeholder {}
+            crop_start -= 5
+
+        max_sequence_length += crop_start
+        text_inputs = self.tokenizer(
+            prompt,
+            max_length=max_sequence_length,
+            padding="max_length",
+            truncation=True,
+            return_tensors="pt",
+            return_length=False,
+            return_overflowing_tokens=False,
+            return_attention_mask=True,
+        )
+        text_input_ids = text_inputs.input_ids.to(device=device)
+        prompt_attention_mask = text_inputs.attention_mask.to(device=device)
+
+        image_embeds = self.image_processor(image, return_tensors="pt").pixel_values.to(device)
+
+        prompt_embeds = self.text_encoder(
+            input_ids=text_input_ids,
+            attention_mask=prompt_attention_mask,
+            pixel_values=image_embeds,
+            output_hidden_states=True,
+        ).hidden_states[-(num_hidden_layers_to_skip + 1)]
+        prompt_embeds = prompt_embeds.to(dtype=dtype)
+
+        image_emb_len = prompt_template.get("image_emb_len", 576)
+        image_emb_start = prompt_template.get("image_emb_start", 5)
+        image_emb_end = prompt_template.get("image_emb_end", 581)
+        double_return_token_id = prompt_template.get("double_return_token_id", 271)
+
+        if crop_start is not None and crop_start > 0:
+            text_crop_start = crop_start - 1 + image_emb_len
+            batch_indices, last_double_return_token_indices = torch.where(text_input_ids == double_return_token_id)
+
+            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([text_input_ids.shape[-1]]))
+                )
+                batch_indices = torch.cat((batch_indices, torch.tensor([0])))
+
+            last_double_return_token_indices = last_double_return_token_indices.reshape(text_input_ids.shape[0], -1)[
+                :, -1
+            ]
+            batch_indices = batch_indices.reshape(text_input_ids.shape[0], -1)[:, -1]
+            assistant_crop_start = last_double_return_token_indices - 1 + image_emb_len - 4
+            assistant_crop_end = last_double_return_token_indices - 1 + image_emb_len
+            attention_mask_assistant_crop_start = last_double_return_token_indices - 4
+            attention_mask_assistant_crop_end = last_double_return_token_indices
+
+            prompt_embed_list = []
+            prompt_attention_mask_list = []
+            image_embed_list = []
+            image_attention_mask_list = []
+
+            for i in range(text_input_ids.shape[0]):
+                prompt_embed_list.append(
+                    torch.cat(
+                        [
+                            prompt_embeds[i, text_crop_start : assistant_crop_start[i].item()],
+                            prompt_embeds[i, assistant_crop_end[i].item() :],
+                        ]
+                    )
+                )
+                prompt_attention_mask_list.append(
+                    torch.cat(
+                        [
+                            prompt_attention_mask[i, crop_start : attention_mask_assistant_crop_start[i].item()],
+                            prompt_attention_mask[i, attention_mask_assistant_crop_end[i].item() :],
+                        ]
+                    )
+                )
+                image_embed_list.append(prompt_embeds[i, image_emb_start:image_emb_end])
+                image_attention_mask_list.append(
+                    torch.ones(image_embed_list[-1].shape[0]).to(prompt_embeds.device).to(prompt_attention_mask.dtype)
+                )
+
+            prompt_embed_list = torch.stack(prompt_embed_list)
+            prompt_attention_mask_list = torch.stack(prompt_attention_mask_list)
+            image_embed_list = torch.stack(image_embed_list)
+            image_attention_mask_list = torch.stack(image_attention_mask_list)
+
+            if 0 < image_embed_interleave < 6:
+                image_embed_list = image_embed_list[:, ::image_embed_interleave, :]
+                image_attention_mask_list = image_attention_mask_list[:, ::image_embed_interleave]
+
+            if not (
+                prompt_embed_list.shape[0] == prompt_attention_mask_list.shape[0]
+                and image_embed_list.shape[0] == image_attention_mask_list.shape[0]
+            ):
+                raise ValueError(
+                    "Input tensors have mismatched batch dimensions."
+                )
+
+            prompt_embeds = torch.cat([image_embed_list, prompt_embed_list], dim=1)
+            prompt_attention_mask = torch.cat([image_attention_mask_list, prompt_attention_mask_list], dim=1)
+
+        return prompt_embeds, prompt_attention_mask
+
+    def _get_clip_prompt_embeds(
+        self,
+        prompt: Union[str, List[str]],
+        num_videos_per_prompt: int = 1,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 77,
+    ) -> torch.Tensor:
+        device = device or self._execution_device
+        dtype = dtype or self.text_encoder_2.dtype
+
+        prompt = [prompt] if isinstance(prompt, str) else prompt
+
+        text_inputs = self.tokenizer_2(
+            prompt,
+            padding="max_length",
+            max_length=max_sequence_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+
+        text_input_ids = text_inputs.input_ids
+        untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids
+        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+            removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, max_sequence_length - 1 : -1])
+            # logger.warning(
+            #     "The following part of your input was truncated because CLIP can only handle sequences up to"
+            #     f" {max_sequence_length} tokens: {removed_text}"
+            # )
+
+        prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False).pooler_output
+        return prompt_embeds
+
+    def encode_prompt(
+        self,
+        image: torch.Tensor,
+        prompt: Union[str, List[str]],
+        prompt_condition: Union[str, List[str], None] = None,
+        prompt_2: Union[str, List[str]] = None,
+        prompt_template: Dict[str, Any] = DEFAULT_PROMPT_TEMPLATE,
+        num_videos_per_prompt: int = 1,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_embeds_condition: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        prompt_attention_mask_condition: Optional[torch.Tensor] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 256,
+        image_embed_interleave: int = 2,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds(
+                image,
+                prompt,
+                prompt_template,
+                num_videos_per_prompt,
+                device=device,
+                dtype=dtype,
+                max_sequence_length=max_sequence_length,
+                image_embed_interleave=image_embed_interleave,
+            )
+
+        if prompt_condition is not None and (prompt_embeds_condition is None or prompt_attention_mask_condition is None):
+            prompt_embeds_condition, prompt_attention_mask_condition = self._get_llama_prompt_embeds(
+                image,
+                prompt_condition,
+                prompt_template,
+                num_videos_per_prompt,
+                device=device,
+                dtype=dtype,
+                max_sequence_length=max_sequence_length,
+                image_embed_interleave=image_embed_interleave,
+            )
+        else:
+            prompt_embeds_condition = prompt_embeds_condition
+            prompt_attention_mask_condition = prompt_attention_mask_condition
+
+        if pooled_prompt_embeds is None:
+            if prompt_2 is None:
+                prompt_2 = prompt
+            pooled_prompt_embeds = self._get_clip_prompt_embeds(
+                prompt,
+                num_videos_per_prompt,
+                device=device,
+                dtype=dtype,
+                max_sequence_length=77,
+            )
+
+        return prompt_embeds, prompt_embeds_condition, pooled_prompt_embeds, prompt_attention_mask, prompt_attention_mask_condition
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        height,
+        width,
+        prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_template=None,
+        true_cfg_scale=1.0,
+        guidance_scale=1.0,
+    ):
+        if height % 16 != 0 or width % 16 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
+
+        if callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt_2 is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+        if prompt_template is not None:
+            if not isinstance(prompt_template, dict):
+                raise ValueError(f"`prompt_template` has to be of type `dict` but is {type(prompt_template)}")
+            if "template" not in prompt_template:
+                raise ValueError(
+                    f"`prompt_template` has to contain a key `template` but only found {prompt_template.keys()}"
+                )
+
+        if true_cfg_scale > 1.0 and guidance_scale > 1.0:
+            logger.warning(
+                "Both `true_cfg_scale` and `guidance_scale` are greater than 1.0. This will result in both "
+                "classifier-free guidance and embedded-guidance to be applied. This is not recommended "
+                "as it may lead to higher memory usage, slower inference and potentially worse results."
+            )
+
+    def prepare_latents(
+        self,
+        image: torch.Tensor,
+        batch_size: int,
+        num_channels_latents: int = 32,
+        height: int = 720,
+        width: int = 1280,
+        num_frames: int = 129,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        image_condition_type: str = "latent_concat",
+    ) -> torch.Tensor:
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
+        latent_height, latent_width = height // self.vae_scale_factor_spatial, width // self.vae_scale_factor_spatial
+        shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width)
+
+        image = image.unsqueeze(2)  # [B, C, 1, H, W]
+        if isinstance(generator, list):
+            image_latents = [
+                retrieve_latents(self.vae.encode(image[i].unsqueeze(0)), generator[i], "argmax")
+                for i in range(batch_size)
+            ]
+        else:
+            image_latents = [retrieve_latents(self.vae.encode(img.unsqueeze(0)), generator, "argmax") for img in image]
+
+        image_latents = torch.cat(image_latents, dim=0).to(dtype) * self.vae_scaling_factor
+        image_latents = image_latents.repeat(1, 1, num_latent_frames, 1, 1)
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device=device, dtype=dtype)
+
+        t = torch.tensor([0.999]).to(device=device)
+        latents = latents * t + image_latents * (1 - t)
+
+        if image_condition_type == "token_replace":
+            image_latents = image_latents[:, :, :1]
+
+        return latents, image_latents
+
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_vae_tiling(self):
+        r"""
+        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+        processing larger images.
+        """
+        self.vae.enable_tiling()
+
+    def disable_vae_tiling(self):
+        r"""
+        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_tiling()
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
+    @property
+    def current_timestep(self):
+        return self._current_timestep
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: PIL.Image.Image,
+        prompt: Union[str, List[str]] = None,
+        prompt_condition: Union[str, List[str], None] = None,
+        prompt_2: Union[str, List[str]] = None,
+        negative_prompt: Union[str, List[str]] = None,
+        negative_prompt_2: Union[str, List[str]] = None,
+        height: int = 720,
+        width: int = 1280,
+        num_frames: int = 129,
+        num_inference_steps: int = 50,
+        sigmas: List[float] = None,
+        true_cfg_scale: float = 1.0,
+        guidance_scale: float = 1.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_embeds_condition: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        prompt_attention_mask_condition: Optional[torch.Tensor] = None,
+        negative_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[
+            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
+        ] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        prompt_template: Dict[str, Any] = DEFAULT_PROMPT_TEMPLATE,
+        max_sequence_length: int = 256,
+        image_embed_interleave: Optional[int] = None,
+        frame_gap: Union[int, None] = None,
+        mixup: bool = False,
+        mixup_num_imgs: Union[int, None] = None,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                will be used instead.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is
+                not greater than `1`).
+            negative_prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+                `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders.
+            height (`int`, defaults to `720`):
+                The height in pixels of the generated image.
+            width (`int`, defaults to `1280`):
+                The width in pixels of the generated image.
+            num_frames (`int`, defaults to `129`):
+                The number of frames in the generated video.
+            num_inference_steps (`int`, defaults to `50`):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            sigmas (`List[float]`, *optional*):
+                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
+                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
+                will be used.
+            true_cfg_scale (`float`, *optional*, defaults to 1.0):
+                When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance.
+            guidance_scale (`float`, defaults to `1.0`):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality. Note that the only available HunyuanVideo model is
+                CFG-distilled, which means that traditional guidance between unconditional and conditional latent is
+                not applied.
+            num_videos_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+                generation deterministic.
+            latents (`torch.Tensor`, *optional*):
+                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor is generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.Tensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+                input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`HunyuanVideoPipelineOutput`] instead of a plain tuple.
+            attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
+                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
+                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
+                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
+                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~HunyuanVideoPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`HunyuanVideoPipelineOutput`] is returned, otherwise a `tuple` is returned
+                where the first element is a list with the generated images and the second element is a list of `bool`s
+                indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content.
+        """
+
+        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
+            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            prompt_embeds,
+            callback_on_step_end_tensor_inputs,
+            prompt_template,
+            true_cfg_scale,
+            guidance_scale,
+        )
+
+        image_condition_type = self.transformer.config.image_condition_type
+        has_neg_prompt = negative_prompt is not None or (
+            negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None
+        )
+        do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
+        image_embed_interleave = (
+            image_embed_interleave
+            if image_embed_interleave is not None
+            else (
+                2 if image_condition_type == "latent_concat" else 4 if image_condition_type == "token_replace" else 1
+            )
+        )
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        # 3. Prepare latent variables
+        vae_dtype = self.vae.dtype
+        image_tensor = self.video_processor.preprocess(image, height, width).to(device, vae_dtype)
+
+        if image_condition_type == "latent_concat":
+            num_channels_latents = (self.transformer.config.in_channels - 1) // 2
+        elif image_condition_type == "token_replace":
+            num_channels_latents = self.transformer.config.in_channels
+
+        latents, image_latents = self.prepare_latents(
+            image_tensor,
+            batch_size * num_videos_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            num_frames if not mixup else num_frames + 4 * mixup_num_imgs,
+            torch.float32,
+            device,
+            generator,
+            latents,
+            image_condition_type,
+        )
+        if image_condition_type == "latent_concat":
+            image_latents[:, :, 1:] = 0
+            mask = image_latents.new_ones(image_latents.shape[0], 1, *image_latents.shape[2:])
+            mask[:, :, 1:] = 0
+
+        # 4. Encode input prompt
+        transformer_dtype = self.transformer.dtype
+        (prompt_embeds, 
+         prompt_embeds_condition, 
+         pooled_prompt_embeds, 
+         prompt_attention_mask, 
+         prompt_attention_mask_condition) = self.encode_prompt(
+            image=image,
+            prompt=prompt,
+            prompt_condition=prompt_condition,
+            prompt_2=prompt_2,
+            prompt_template=prompt_template,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            prompt_embeds_condition=prompt_embeds_condition,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            prompt_attention_mask_condition=prompt_attention_mask_condition,
+            device=device,
+            max_sequence_length=max_sequence_length,
+            image_embed_interleave=image_embed_interleave,
+        )
+        prompt_embeds = prompt_embeds.to(transformer_dtype)
+        prompt_attention_mask = prompt_attention_mask.to(transformer_dtype)
+        pooled_prompt_embeds = pooled_prompt_embeds.to(transformer_dtype)
+
+        if do_true_cfg:
+            black_image = PIL.Image.new("RGB", (width, height), 0)
+            negative_prompt_embeds, negative_pooled_prompt_embeds, negative_prompt_attention_mask = self.encode_prompt(
+                image=black_image,
+                prompt=negative_prompt,
+                prompt_2=negative_prompt_2,
+                prompt_template=prompt_template,
+                num_videos_per_prompt=num_videos_per_prompt,
+                prompt_embeds=negative_prompt_embeds,
+                pooled_prompt_embeds=negative_pooled_prompt_embeds,
+                prompt_attention_mask=negative_prompt_attention_mask,
+                device=device,
+                max_sequence_length=max_sequence_length,
+            )
+            negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype)
+            negative_prompt_attention_mask = negative_prompt_attention_mask.to(transformer_dtype)
+            negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.to(transformer_dtype)
+
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, sigmas=sigmas)
+
+        # 6. Prepare guidance condition
+        guidance = None
+        if self.transformer.config.guidance_embeds:
+            guidance = (
+                torch.tensor([guidance_scale] * latents.shape[0], dtype=transformer_dtype, device=device) * 1000.0
+            )
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                self._current_timestep = t
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                if image_condition_type == "latent_concat":
+                    latent_model_input = torch.cat([latents, image_latents, mask], dim=1).to(transformer_dtype)
+                elif image_condition_type == "token_replace":
+                    latent_model_input = torch.cat([image_latents, latents[:, :, 1:]], dim=2).to(transformer_dtype)
+
+                noise_pred = self.transformer(
+                    hidden_states=latent_model_input,
+                    timestep=timestep,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_hidden_states_condition=prompt_embeds_condition,
+                    encoder_attention_mask=prompt_attention_mask,
+                    encoder_attention_mask_condition=prompt_attention_mask_condition,
+                    pooled_projections=pooled_prompt_embeds,
+                    guidance=guidance,
+                    attention_kwargs=attention_kwargs,
+                    return_dict=False,
+                    frame_gap=int(frame_gap / 4) if frame_gap is not None else frame_gap,
+                )[0]
+
+                if do_true_cfg:
+                    neg_noise_pred = self.transformer(
+                        hidden_states=latent_model_input,
+                        timestep=timestep,
+                        encoder_hidden_states=negative_prompt_embeds,
+                        encoder_attention_mask=negative_prompt_attention_mask,
+                        pooled_projections=negative_pooled_prompt_embeds,
+                        guidance=guidance,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+                    noise_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                if image_condition_type == "latent_concat":
+                    latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+                elif image_condition_type == "token_replace":
+                    latents = latents = self.scheduler.step(
+                        noise_pred[:, :, 1:], t, latents[:, :, 1:], return_dict=False
+                    )[0]
+                    latents = torch.cat([image_latents, latents], dim=2)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        self._current_timestep = None
+
+        if not mixup:
+            generated_img_frame_start = image_latents.shape[2]
+            latents = latents[:, :, generated_img_frame_start:]
+        if not output_type == "latent":
+            latents = latents.to(self.vae.dtype) / self.vae_scaling_factor
+            video = self.vae.decode(latents, return_dict=False)[0]
+            if image_condition_type == "latent_concat":
+                video = video[:, :, 4:, :, :]
+            video = self.video_processor.postprocess_video(video, output_type=output_type)
+            if mixup:
+                single_generated_decoded = self.vae.decode(latents[:, :, -1:], return_dict=False)[0]
+                single_generated_decoded = self.video_processor.postprocess_video(single_generated_decoded, output_type=output_type)
+                video = torch.cat([single_generated_decoded, video], dim=1)
+        else:
+            if image_condition_type == "latent_concat":
+                video = latents[:, :, 1:, :, :]
+            else:
+                video = latents
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return HunyuanVideoPipelineOutput(frames=video)

requirements.txt

@@ -0,0 +1,9 @@
+torch==2.4.1
+torchvision==0.19.1
+diffusers==0.33.1
+transformers==4.45.0
+flash-attn==2.7.3
+gradio
+omegaconf
+peft
+opencv-python