Hunyuan Video Framepack (#11428)

* add transformer

* add pipeline

* fixes

* make fix-copies

* update

* add flux mu shift

* update example snippet

* debug

* cleanup

* batch_size=1 optimization

* add pipeline test

* fix for model cpu offloading'

* add last_image support; credits: https://github.com/lllyasviel/FramePack/pull/167

* update example with flf2v

* update penguin url

* fix test

* address review comment: https://github.com/huggingface/diffusers/pull/11428#discussion_r2071032371

* address review comment: https://github.com/huggingface/diffusers/pull/11428#discussion_r2071087689



* Update src/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_framepack.py

---------
Co-authored-by: Linoy Tsaban <57615435+linoytsaban@users.noreply.github.com>

docs/source/en/api/pipelines/hunyuan_video.md

@@ -52,6 +52,7 @@ The following models are available for the image-to-video pipeline:
 | [`Skywork/SkyReels-V1-Hunyuan-I2V`](https://huggingface.co/Skywork/SkyReels-V1-Hunyuan-I2V) | Skywork's custom finetune of HunyuanVideo (de-distilled). Performs best with `97x544x960` resolution. Performs best at `97x544x960` resolution, `guidance_scale=1.0`, `true_cfg_scale=6.0` and a negative prompt. |
 | [`hunyuanvideo-community/HunyuanVideo-I2V-33ch`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo 33-channel I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20). |
 | [`hunyuanvideo-community/HunyuanVideo-I2V`](https://huggingface.co/hunyuanvideo-community/HunyuanVideo-I2V) | Tecent's official HunyuanVideo 16-channel I2V model. Performs best at resolutions of 480, 720, 960, 1280. A higher `shift` value when initializing the scheduler is recommended (good values are between 7 and 20) |
+- [`lllyasviel/FramePackI2V_HY`](https://huggingface.co/lllyasviel/FramePackI2V_HY) | lllyasviel's paper introducing a new technique for long-context video generation called [Framepack](https://arxiv.org/abs/2504.12626). |
 
 ## Quantization
 

src/diffusers/__init__.py

@@ -175,6 +175,7 @@ else:
             "HunyuanDiT2DControlNetModel",
             "HunyuanDiT2DModel",
             "HunyuanDiT2DMultiControlNetModel",
+            "HunyuanVideoFramepackTransformer3DModel",
             "HunyuanVideoTransformer3DModel",
             "I2VGenXLUNet",
             "Kandinsky3UNet",
@@ -376,6 +377,7 @@ else:
             "HunyuanDiTPAGPipeline",
             "HunyuanDiTPipeline",
             "HunyuanSkyreelsImageToVideoPipeline",
+            "HunyuanVideoFramepackPipeline",
             "HunyuanVideoImageToVideoPipeline",
             "HunyuanVideoPipeline",
             "I2VGenXLPipeline",
@@ -770,6 +772,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             HunyuanDiT2DControlNetModel,
             HunyuanDiT2DModel,
             HunyuanDiT2DMultiControlNetModel,
+            HunyuanVideoFramepackTransformer3DModel,
             HunyuanVideoTransformer3DModel,
             I2VGenXLUNet,
             Kandinsky3UNet,
@@ -950,6 +953,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             HunyuanDiTPAGPipeline,
             HunyuanDiTPipeline,
             HunyuanSkyreelsImageToVideoPipeline,
+            HunyuanVideoFramepackPipeline,
             HunyuanVideoImageToVideoPipeline,
             HunyuanVideoPipeline,
             I2VGenXLPipeline,

src/diffusers/models/__init__.py

@@ -79,6 +79,7 @@ if is_torch_available():
     _import_structure["transformers.transformer_flux"] = ["FluxTransformer2DModel"]
     _import_structure["transformers.transformer_hidream_image"] = ["HiDreamImageTransformer2DModel"]
     _import_structure["transformers.transformer_hunyuan_video"] = ["HunyuanVideoTransformer3DModel"]
+    _import_structure["transformers.transformer_hunyuan_video_framepack"] = ["HunyuanVideoFramepackTransformer3DModel"]
     _import_structure["transformers.transformer_ltx"] = ["LTXVideoTransformer3DModel"]
     _import_structure["transformers.transformer_lumina2"] = ["Lumina2Transformer2DModel"]
     _import_structure["transformers.transformer_mochi"] = ["MochiTransformer3DModel"]
@@ -156,6 +157,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             FluxTransformer2DModel,
             HiDreamImageTransformer2DModel,
             HunyuanDiT2DModel,
+            HunyuanVideoFramepackTransformer3DModel,
             HunyuanVideoTransformer3DModel,
             LatteTransformer3DModel,
             LTXVideoTransformer3DModel,

src/diffusers/models/transformers/__init__.py

@@ -23,6 +23,7 @@ if is_torch_available():
     from .transformer_flux import FluxTransformer2DModel
     from .transformer_hidream_image import HiDreamImageTransformer2DModel
     from .transformer_hunyuan_video import HunyuanVideoTransformer3DModel
+    from .transformer_hunyuan_video_framepack import HunyuanVideoFramepackTransformer3DModel
     from .transformer_ltx import LTXVideoTransformer3DModel
     from .transformer_lumina2 import Lumina2Transformer2DModel
     from .transformer_mochi import MochiTransformer3DModel

src/diffusers/models/transformers/transformer_hunyuan_video_framepack.py

+# Copyright 2025 The Framepack Team, 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.
+
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalModelMixin, PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, get_logger, scale_lora_layers, unscale_lora_layers
+from ..cache_utils import CacheMixin
+from ..embeddings import get_1d_rotary_pos_embed
+from ..modeling_outputs import Transformer2DModelOutput
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormContinuous
+from .transformer_hunyuan_video import (
+    HunyuanVideoConditionEmbedding,
+    HunyuanVideoPatchEmbed,
+    HunyuanVideoSingleTransformerBlock,
+    HunyuanVideoTokenRefiner,
+    HunyuanVideoTransformerBlock,
+)
+
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class HunyuanVideoFramepackRotaryPosEmbed(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, frame_indices: torch.Tensor, height: int, width: int, device: torch.device):
+        height = height // self.patch_size
+        width = width // self.patch_size
+        grid = torch.meshgrid(
+            frame_indices.to(device=device, dtype=torch.float32),
+            torch.arange(0, height, device=device, dtype=torch.float32),
+            torch.arange(0, width, device=device, dtype=torch.float32),
+            indexing="ij",
+        )  # 3 * [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 FramepackClipVisionProjection(nn.Module):
+    def __init__(self, in_channels: int, out_channels: int):
+        super().__init__()
+        self.up = nn.Linear(in_channels, out_channels * 3)
+        self.down = nn.Linear(out_channels * 3, out_channels)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.up(hidden_states)
+        hidden_states = F.silu(hidden_states)
+        hidden_states = self.down(hidden_states)
+        return hidden_states
+
+
+class HunyuanVideoHistoryPatchEmbed(nn.Module):
+    def __init__(self, in_channels: int, inner_dim: int):
+        super().__init__()
+        self.proj = nn.Conv3d(in_channels, inner_dim, kernel_size=(1, 2, 2), stride=(1, 2, 2))
+        self.proj_2x = nn.Conv3d(in_channels, inner_dim, kernel_size=(2, 4, 4), stride=(2, 4, 4))
+        self.proj_4x = nn.Conv3d(in_channels, inner_dim, kernel_size=(4, 8, 8), stride=(4, 8, 8))
+
+    def forward(
+        self,
+        latents_clean: Optional[torch.Tensor] = None,
+        latents_clean_2x: Optional[torch.Tensor] = None,
+        latents_clean_4x: Optional[torch.Tensor] = None,
+    ):
+        if latents_clean is not None:
+            latents_clean = self.proj(latents_clean)
+            latents_clean = latents_clean.flatten(2).transpose(1, 2)
+        if latents_clean_2x is not None:
+            latents_clean_2x = _pad_for_3d_conv(latents_clean_2x, (2, 4, 4))
+            latents_clean_2x = self.proj_2x(latents_clean_2x)
+            latents_clean_2x = latents_clean_2x.flatten(2).transpose(1, 2)
+        if latents_clean_4x is not None:
+            latents_clean_4x = _pad_for_3d_conv(latents_clean_4x, (4, 8, 8))
+            latents_clean_4x = self.proj_4x(latents_clean_4x)
+            latents_clean_4x = latents_clean_4x.flatten(2).transpose(1, 2)
+        return latents_clean, latents_clean_2x, latents_clean_4x
+
+
+class HunyuanVideoFramepackTransformer3DModel(
+    ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin, CacheMixin
+):
+    _supports_gradient_checkpointing = True
+    _skip_layerwise_casting_patterns = ["x_embedder", "context_embedder", "norm"]
+    _no_split_modules = [
+        "HunyuanVideoTransformerBlock",
+        "HunyuanVideoSingleTransformerBlock",
+        "HunyuanVideoHistoryPatchEmbed",
+        "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,
+        has_image_proj: int = False,
+        image_proj_dim: int = 1152,
+        has_clean_x_embedder: int = False,
+    ) -> None:
+        super().__init__()
+
+        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 = HunyuanVideoFramepackRotaryPosEmbed(patch_size, patch_size_t, rope_axes_dim, rope_theta)
+
+        # 3. Dual stream transformer blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                HunyuanVideoTransformerBlock(
+                    num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        # 4. Single stream transformer blocks
+        self.single_transformer_blocks = nn.ModuleList(
+            [
+                HunyuanVideoSingleTransformerBlock(
+                    num_attention_heads, attention_head_dim, mlp_ratio=mlp_ratio, qk_norm=qk_norm
+                )
+                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)
+
+        # Framepack specific modules
+        self.image_projection = FramepackClipVisionProjection(image_proj_dim, inner_dim) if has_image_proj else None
+
+        self.clean_x_embedder = None
+        if has_clean_x_embedder:
+            self.clean_x_embedder = HunyuanVideoHistoryPatchEmbed(in_channels, inner_dim)
+
+        self.use_gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        timestep: torch.LongTensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_attention_mask: torch.Tensor,
+        pooled_projections: torch.Tensor,
+        image_embeds: torch.Tensor,
+        indices_latents: torch.Tensor,
+        guidance: Optional[torch.Tensor] = None,
+        latents_clean: Optional[torch.Tensor] = None,
+        indices_latents_clean: Optional[torch.Tensor] = None,
+        latents_history_2x: Optional[torch.Tensor] = None,
+        indices_latents_history_2x: Optional[torch.Tensor] = None,
+        latents_history_4x: Optional[torch.Tensor] = None,
+        indices_latents_history_4x: Optional[torch.Tensor] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ):
+        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
+        original_context_length = post_patch_num_frames * post_patch_height * post_patch_width
+
+        if indices_latents is None:
+            indices_latents = torch.arange(0, num_frames).unsqueeze(0).expand(batch_size, -1)
+
+        hidden_states = self.x_embedder(hidden_states)
+        image_rotary_emb = self.rope(
+            frame_indices=indices_latents, height=height, width=width, device=hidden_states.device
+        )
+
+        latents_clean, latents_history_2x, latents_history_4x = self.clean_x_embedder(
+            latents_clean, latents_history_2x, latents_history_4x
+        )
+
+        if latents_clean is not None and indices_latents_clean is not None:
+            image_rotary_emb_clean = self.rope(
+                frame_indices=indices_latents_clean, height=height, width=width, device=hidden_states.device
+            )
+        if latents_history_2x is not None and indices_latents_history_2x is not None:
+            image_rotary_emb_history_2x = self.rope(
+                frame_indices=indices_latents_history_2x, height=height, width=width, device=hidden_states.device
+            )
+        if latents_history_4x is not None and indices_latents_history_4x is not None:
+            image_rotary_emb_history_4x = self.rope(
+                frame_indices=indices_latents_history_4x, height=height, width=width, device=hidden_states.device
+            )
+
+        hidden_states, image_rotary_emb = self._pack_history_states(
+            hidden_states,
+            latents_clean,
+            latents_history_2x,
+            latents_history_4x,
+            image_rotary_emb,
+            image_rotary_emb_clean,
+            image_rotary_emb_history_2x,
+            image_rotary_emb_history_4x,
+            post_patch_height,
+            post_patch_width,
+        )
+
+        temb, _ = self.time_text_embed(timestep, pooled_projections, guidance)
+        encoder_hidden_states = self.context_embedder(encoder_hidden_states, timestep, encoder_attention_mask)
+
+        encoder_hidden_states_image = self.image_projection(image_embeds)
+        attention_mask_image = encoder_attention_mask.new_ones((batch_size, encoder_hidden_states_image.shape[1]))
+
+        # must cat before (not after) encoder_hidden_states, due to attn masking
+        encoder_hidden_states = torch.cat([encoder_hidden_states_image, encoder_hidden_states], dim=1)
+        encoder_attention_mask = torch.cat([attention_mask_image, encoder_attention_mask], dim=1)
+
+        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
+
+        if batch_size == 1:
+            encoder_hidden_states = encoder_hidden_states[:, : effective_condition_sequence_length[0]]
+            attention_mask = None
+        else:
+            for i in range(batch_size):
+                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)
+
+        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
+                )
+
+            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
+                )
+
+        else:
+            for block in self.transformer_blocks:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states, encoder_hidden_states, temb, attention_mask, image_rotary_emb
+                )
+
+            for block in self.single_transformer_blocks:
+                hidden_states, encoder_hidden_states = block(
+                    hidden_states, encoder_hidden_states, temb, attention_mask, image_rotary_emb
+                )
+
+        hidden_states = hidden_states[:, -original_context_length:]
+        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)
+
+    def _pack_history_states(
+        self,
+        hidden_states: torch.Tensor,
+        latents_clean: Optional[torch.Tensor] = None,
+        latents_history_2x: Optional[torch.Tensor] = None,
+        latents_history_4x: Optional[torch.Tensor] = None,
+        image_rotary_emb: Tuple[torch.Tensor, torch.Tensor] = None,
+        image_rotary_emb_clean: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        image_rotary_emb_history_2x: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        image_rotary_emb_history_4x: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        height: int = None,
+        width: int = None,
+    ):
+        image_rotary_emb = list(image_rotary_emb)  # convert tuple to list for in-place modification
+
+        if latents_clean is not None and image_rotary_emb_clean is not None:
+            hidden_states = torch.cat([latents_clean, hidden_states], dim=1)
+            image_rotary_emb[0] = torch.cat([image_rotary_emb_clean[0], image_rotary_emb[0]], dim=0)
+            image_rotary_emb[1] = torch.cat([image_rotary_emb_clean[1], image_rotary_emb[1]], dim=0)
+
+        if latents_history_2x is not None and image_rotary_emb_history_2x is not None:
+            hidden_states = torch.cat([latents_history_2x, hidden_states], dim=1)
+            image_rotary_emb_history_2x = self._pad_rotary_emb(image_rotary_emb_history_2x, height, width, (2, 2, 2))
+            image_rotary_emb[0] = torch.cat([image_rotary_emb_history_2x[0], image_rotary_emb[0]], dim=0)
+            image_rotary_emb[1] = torch.cat([image_rotary_emb_history_2x[1], image_rotary_emb[1]], dim=0)
+
+        if latents_history_4x is not None and image_rotary_emb_history_4x is not None:
+            hidden_states = torch.cat([latents_history_4x, hidden_states], dim=1)
+            image_rotary_emb_history_4x = self._pad_rotary_emb(image_rotary_emb_history_4x, height, width, (4, 4, 4))
+            image_rotary_emb[0] = torch.cat([image_rotary_emb_history_4x[0], image_rotary_emb[0]], dim=0)
+            image_rotary_emb[1] = torch.cat([image_rotary_emb_history_4x[1], image_rotary_emb[1]], dim=0)
+
+        return hidden_states, tuple(image_rotary_emb)
+
+    def _pad_rotary_emb(
+        self,
+        image_rotary_emb: Tuple[torch.Tensor],
+        height: int,
+        width: int,
+        kernel_size: Tuple[int, int, int],
+    ):
+        # freqs_cos, freqs_sin have shape [W * H * T, D / 2], where D is attention head dim
+        freqs_cos, freqs_sin = image_rotary_emb
+        freqs_cos = freqs_cos.unsqueeze(0).permute(0, 2, 1).unflatten(2, (-1, height, width))
+        freqs_sin = freqs_sin.unsqueeze(0).permute(0, 2, 1).unflatten(2, (-1, height, width))
+        freqs_cos = _pad_for_3d_conv(freqs_cos, kernel_size)
+        freqs_sin = _pad_for_3d_conv(freqs_sin, kernel_size)
+        freqs_cos = _center_down_sample_3d(freqs_cos, kernel_size)
+        freqs_sin = _center_down_sample_3d(freqs_sin, kernel_size)
+        freqs_cos = freqs_cos.flatten(2).permute(0, 2, 1).squeeze(0)
+        freqs_sin = freqs_sin.flatten(2).permute(0, 2, 1).squeeze(0)
+        return freqs_cos, freqs_sin
+
+
+def _pad_for_3d_conv(x, kernel_size):
+    if isinstance(x, (tuple, list)):
+        return tuple(_pad_for_3d_conv(i, kernel_size) for i in x)
+    b, c, t, h, w = x.shape
+    pt, ph, pw = kernel_size
+    pad_t = (pt - (t % pt)) % pt
+    pad_h = (ph - (h % ph)) % ph
+    pad_w = (pw - (w % pw)) % pw
+    return torch.nn.functional.pad(x, (0, pad_w, 0, pad_h, 0, pad_t), mode="replicate")
+
+
+def _center_down_sample_3d(x, kernel_size):
+    if isinstance(x, (tuple, list)):
+        return tuple(_center_down_sample_3d(i, kernel_size) for i in x)
+    return torch.nn.functional.avg_pool3d(x, kernel_size, stride=kernel_size)

src/diffusers/pipelines/__init__.py

@@ -227,6 +227,7 @@ else:
         "HunyuanVideoPipeline",
         "HunyuanSkyreelsImageToVideoPipeline",
         "HunyuanVideoImageToVideoPipeline",
+        "HunyuanVideoFramepackPipeline",
     ]
     _import_structure["kandinsky"] = [
         "KandinskyCombinedPipeline",
@@ -589,6 +590,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .hidream_image import HiDreamImagePipeline
         from .hunyuan_video import (
             HunyuanSkyreelsImageToVideoPipeline,
+            HunyuanVideoFramepackPipeline,
             HunyuanVideoImageToVideoPipeline,
             HunyuanVideoPipeline,
         )

src/diffusers/pipelines/hunyuan_video/__init__.py

@@ -24,6 +24,7 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["pipeline_hunyuan_skyreels_image2video"] = ["HunyuanSkyreelsImageToVideoPipeline"]
     _import_structure["pipeline_hunyuan_video"] = ["HunyuanVideoPipeline"]
+    _import_structure["pipeline_hunyuan_video_framepack"] = ["HunyuanVideoFramepackPipeline"]
     _import_structure["pipeline_hunyuan_video_image2video"] = ["HunyuanVideoImageToVideoPipeline"]
 
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
@@ -36,6 +37,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     else:
         from .pipeline_hunyuan_skyreels_image2video import HunyuanSkyreelsImageToVideoPipeline
         from .pipeline_hunyuan_video import HunyuanVideoPipeline
+        from .pipeline_hunyuan_video_framepack import HunyuanVideoFramepackPipeline
         from .pipeline_hunyuan_video_image2video import HunyuanVideoImageToVideoPipeline
 
 else:

src/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video_framepack.py

+# Copyright 2025 The Framepack Team, 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.
+
+import inspect
+import math
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from transformers import (
+    CLIPTextModel,
+    CLIPTokenizer,
+    LlamaModel,
+    LlamaTokenizerFast,
+    SiglipImageProcessor,
+    SiglipVisionModel,
+)
+
+from ...callbacks import MultiPipelineCallbacks, PipelineCallback
+from ...image_processor import PipelineImageInput
+from ...loaders import HunyuanVideoLoraLoaderMixin
+from ...models import AutoencoderKLHunyuanVideo, HunyuanVideoFramepackTransformer3DModel
+from ...schedulers import FlowMatchEulerDiscreteScheduler
+from ...utils import is_torch_xla_available, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ...video_processor import VideoProcessor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import HunyuanVideoFramepackPipelineOutput
+
+
+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
+
+
+# TODO(yiyi): We can pack the checkpoints nicely with modular loader
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ##### Image-to-Video
+
+        ```python
+        >>> import torch
+        >>> from diffusers import HunyuanVideoFramepackPipeline, HunyuanVideoFramepackTransformer3DModel
+        >>> from diffusers.utils import export_to_video, load_image
+        >>> from transformers import SiglipImageProcessor, SiglipVisionModel
+
+        >>> transformer = HunyuanVideoFramepackTransformer3DModel.from_pretrained(
+        ...     "lllyasviel/FramePackI2V_HY", torch_dtype=torch.bfloat16
+        ... )
+        >>> feature_extractor = SiglipImageProcessor.from_pretrained(
+        ...     "lllyasviel/flux_redux_bfl", subfolder="feature_extractor"
+        ... )
+        >>> image_encoder = SiglipVisionModel.from_pretrained(
+        ...     "lllyasviel/flux_redux_bfl", subfolder="image_encoder", torch_dtype=torch.float16
+        ... )
+        >>> pipe = HunyuanVideoFramepackPipeline.from_pretrained(
+        ...     "hunyuanvideo-community/HunyuanVideo",
+        ...     transformer=transformer,
+        ...     feature_extractor=feature_extractor,
+        ...     image_encoder=image_encoder,
+        ...     torch_dtype=torch.float16,
+        ... )
+        >>> pipe.vae.enable_tiling()
+        >>> pipe.to("cuda")
+
+        >>> image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/penguin.png"
+        ... )
+        >>> output = pipe(
+        ...     image=image,
+        ...     prompt="A penguin dancing in the snow",
+        ...     height=832,
+        ...     width=480,
+        ...     num_frames=91,
+        ...     num_inference_steps=30,
+        ...     guidance_scale=9.0,
+        ...     generator=torch.Generator().manual_seed(0),
+        ... ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=30)
+        ```
+
+        ##### First and Last Image-to-Video
+
+        ```python
+        >>> import torch
+        >>> from diffusers import HunyuanVideoFramepackPipeline, HunyuanVideoFramepackTransformer3DModel
+        >>> from diffusers.utils import export_to_video, load_image
+        >>> from transformers import SiglipImageProcessor, SiglipVisionModel
+
+        >>> transformer = HunyuanVideoFramepackTransformer3DModel.from_pretrained(
+        ...     "lllyasviel/FramePackI2V_HY", torch_dtype=torch.bfloat16
+        ... )
+        >>> feature_extractor = SiglipImageProcessor.from_pretrained(
+        ...     "lllyasviel/flux_redux_bfl", subfolder="feature_extractor"
+        ... )
+        >>> image_encoder = SiglipVisionModel.from_pretrained(
+        ...     "lllyasviel/flux_redux_bfl", subfolder="image_encoder", torch_dtype=torch.float16
+        ... )
+        >>> pipe = HunyuanVideoFramepackPipeline.from_pretrained(
+        ...     "hunyuanvideo-community/HunyuanVideo",
+        ...     transformer=transformer,
+        ...     feature_extractor=feature_extractor,
+        ...     image_encoder=image_encoder,
+        ...     torch_dtype=torch.float16,
+        ... )
+        >>> pipe.to("cuda")
+
+        >>> prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective."
+        >>> first_image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png"
+        ... )
+        >>> last_image = load_image(
+        ...     "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png"
+        ... )
+        >>> output = pipe(
+        ...     image=first_image,
+        ...     last_image=last_image,
+        ...     prompt=prompt,
+        ...     height=512,
+        ...     width=512,
+        ...     num_frames=91,
+        ...     num_inference_steps=30,
+        ...     guidance_scale=9.0,
+        ...     generator=torch.Generator().manual_seed(0),
+        ... ).frames[0]
+        >>> export_to_video(output, "output.mp4", fps=30)
+        ```
+"""
+
+
+DEFAULT_PROMPT_TEMPLATE = {
+    "template": (
+        "<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: "
+        "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|>"
+        "<|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>"
+    ),
+    "crop_start": 95,
+}
+
+
+# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
+def calculate_shift(
+    image_seq_len,
+    base_seq_len: int = 256,
+    max_seq_len: int = 4096,
+    base_shift: float = 0.5,
+    max_shift: float = 1.15,
+):
+    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+    b = base_shift - m * base_seq_len
+    mu = image_seq_len * m + b
+    return mu
+
+
+# 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
+
+
+class HunyuanVideoFramepackPipeline(DiffusionPipeline, HunyuanVideoLoraLoaderMixin):
+    r"""
+    Pipeline for text-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 ([`LlamaModel`]):
+            [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->image_encoder->transformer->vae"
+    _callback_tensor_inputs = ["latents", "prompt_embeds"]
+
+    def __init__(
+        self,
+        text_encoder: LlamaModel,
+        tokenizer: LlamaTokenizerFast,
+        transformer: HunyuanVideoFramepackTransformer3DModel,
+        vae: AutoencoderKLHunyuanVideo,
+        scheduler: FlowMatchEulerDiscreteScheduler,
+        text_encoder_2: CLIPTextModel,
+        tokenizer_2: CLIPTokenizer,
+        image_encoder: SiglipVisionModel,
+        feature_extractor: SiglipImageProcessor,
+    ):
+        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_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+        )
+
+        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)
+
+    # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline._get_llama_prompt_embeds
+    def _get_llama_prompt_embeds(
+        self,
+        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,
+    ) -> 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
+        batch_size = len(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 <|eot_id|> token and placeholder {}
+            crop_start -= 2
+
+        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)
+
+        prompt_embeds = self.text_encoder(
+            input_ids=text_input_ids,
+            attention_mask=prompt_attention_mask,
+            output_hidden_states=True,
+        ).hidden_states[-(num_hidden_layers_to_skip + 1)]
+        prompt_embeds = prompt_embeds.to(dtype=dtype)
+
+        if crop_start is not None and crop_start > 0:
+            prompt_embeds = prompt_embeds[:, crop_start:]
+            prompt_attention_mask = prompt_attention_mask[:, crop_start:]
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        _, seq_len, _ = prompt_embeds.shape
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+        prompt_attention_mask = prompt_attention_mask.repeat(1, num_videos_per_prompt)
+        prompt_attention_mask = prompt_attention_mask.view(batch_size * num_videos_per_prompt, seq_len)
+
+        return prompt_embeds, prompt_attention_mask
+
+    # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline._get_clip_prompt_embeds
+    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
+        batch_size = len(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
+
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt)
+        prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, -1)
+
+        return prompt_embeds
+
+    # Copied from diffusers.pipelines.hunyuan_video.pipeline_hunyuan_video.HunyuanVideoPipeline.encode_prompt
+    def encode_prompt(
+        self,
+        prompt: Union[str, List[str]],
+        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,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: Optional[torch.Tensor] = None,
+        device: Optional[torch.device] = None,
+        dtype: Optional[torch.dtype] = None,
+        max_sequence_length: int = 256,
+    ):
+        if prompt_embeds is None:
+            prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds(
+                prompt,
+                prompt_template,
+                num_videos_per_prompt,
+                device=device,
+                dtype=dtype,
+                max_sequence_length=max_sequence_length,
+            )
+
+        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, pooled_prompt_embeds, prompt_attention_mask
+
+    def encode_image(
+        self, image: torch.Tensor, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None
+    ):
+        device = device or self._execution_device
+        image = (image + 1) / 2.0  # [-1, 1] -> [0, 1]
+        image = self.feature_extractor(images=image, return_tensors="pt", do_rescale=False).to(
+            device=device, dtype=self.image_encoder.dtype
+        )
+        image_embeds = self.image_encoder(**image).last_hidden_state
+        return image_embeds.to(dtype=dtype)
+
+    def check_inputs(
+        self,
+        prompt,
+        prompt_2,
+        height,
+        width,
+        prompt_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+        prompt_template=None,
+    ):
+        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()}"
+                )
+
+    def prepare_latents(
+        self,
+        batch_size: int = 1,
+        num_channels_latents: int = 16,
+        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,
+    ) -> torch.Tensor:
+        if latents is not None:
+            return latents.to(device=device, dtype=dtype)
+        shape = (
+            batch_size,
+            num_channels_latents,
+            (num_frames - 1) // self.vae_scale_factor_temporal + 1,
+            int(height) // self.vae_scale_factor_spatial,
+            int(width) // self.vae_scale_factor_spatial,
+        )
+        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."
+            )
+        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        return latents
+
+    def prepare_image_latents(
+        self,
+        image: torch.Tensor,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        device = device or self._execution_device
+        if latents is None:
+            image = image.unsqueeze(2).to(device=device, dtype=self.vae.dtype)
+            latents = self.vae.encode(image).latent_dist.sample(generator=generator)
+            latents = latents * self.vae.config.scaling_factor
+        return latents.to(device=device, dtype=dtype)
+
+    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: PipelineImageInput,
+        last_image: Optional[PipelineImageInput] = None,
+        prompt: Union[str, List[str]] = 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,
+        latent_window_size: int = 9,
+        num_inference_steps: int = 50,
+        sigmas: List[float] = None,
+        true_cfg_scale: float = 1.0,
+        guidance_scale: float = 6.0,
+        num_videos_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        image_latents: Optional[torch.Tensor] = None,
+        last_image_latents: Optional[torch.Tensor] = None,
+        prompt_embeds: Optional[torch.Tensor] = None,
+        pooled_prompt_embeds: Optional[torch.Tensor] = None,
+        prompt_attention_mask: 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,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`):
+                The image to be used as the starting point for the video generation.
+            last_image (`PIL.Image.Image` or `np.ndarray` or `torch.Tensor`, *optional*):
+                The optional last image to be used as the ending point for the video generation. This is useful for
+                generating transitions between two images.
+            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 `6.0`):
+                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.
+            image_latents (`torch.Tensor`, *optional*):
+                Pre-encoded image latents. If not provided, the image will be encoded using the VAE.
+            last_image_latents (`torch.Tensor`, *optional*):
+                Pre-encoded last image latents. If not provided, the last image will be encoded using the VAE.
+            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 [`HunyuanVideoFramepackPipelineOutput`] 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:
+            [`~HunyuanVideoFramepackPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`HunyuanVideoFramepackPipelineOutput`] 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,
+        )
+
+        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
+
+        self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
+        self._current_timestep = None
+        self._interrupt = False
+
+        device = self._execution_device
+        transformer_dtype = self.transformer.dtype
+        vae_dtype = self.vae.dtype
+
+        # 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. Encode input prompt
+        transformer_dtype = self.transformer.dtype
+        prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_template=prompt_template,
+            num_videos_per_prompt=num_videos_per_prompt,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            prompt_attention_mask=prompt_attention_mask,
+            device=device,
+            max_sequence_length=max_sequence_length,
+        )
+        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:
+            negative_prompt_embeds, negative_pooled_prompt_embeds, negative_prompt_attention_mask = self.encode_prompt(
+                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)
+
+        # 4. Prepare image
+        image = self.video_processor.preprocess(image, height, width)
+        image_embeds = self.encode_image(image, device=device).to(transformer_dtype)
+        if last_image is not None:
+            # Credits: https://github.com/lllyasviel/FramePack/pull/167
+            # Users can modify the weighting strategy applied here
+            last_image = self.video_processor.preprocess(last_image, height, width)
+            last_image_embeds = self.encode_image(last_image, device=device).to(transformer_dtype)
+            last_image_embeds = (image_embeds + last_image_embeds) / 2
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels
+        window_num_frames = (latent_window_size - 1) * self.vae_scale_factor_temporal + 1
+        num_latent_sections = max(1, (num_frames + window_num_frames - 1) // window_num_frames)
+        # Specific to the released checkpoint: https://huggingface.co/lllyasviel/FramePackI2V_HY
+        # TODO: find a more generic way in future if there are more checkpoints
+        history_sizes = [1, 2, 16]
+        history_latents = torch.zeros(
+            batch_size,
+            num_channels_latents,
+            sum(history_sizes),
+            height // self.vae_scale_factor_spatial,
+            width // self.vae_scale_factor_spatial,
+            device=device,
+            dtype=torch.float32,
+        )
+        history_video = None
+        total_generated_latent_frames = 0
+
+        image_latents = self.prepare_image_latents(
+            image, dtype=torch.float32, device=device, generator=generator, latents=image_latents
+        )
+        if last_image is not None:
+            last_image_latents = self.prepare_image_latents(
+                last_image, dtype=torch.float32, device=device, generator=generator
+            )
+
+        latent_paddings = list(reversed(range(num_latent_sections)))
+        if num_latent_sections > 4:
+            latent_paddings = [3] + [2] * (num_latent_sections - 3) + [1, 0]
+
+        # 6. Prepare guidance condition
+        guidance = torch.tensor([guidance_scale] * batch_size, dtype=transformer_dtype, device=device) * 1000.0
+
+        # 7. Denoising loop
+        for k in range(num_latent_sections):
+            is_first_section = k == 0
+            is_last_section = k == num_latent_sections - 1
+            latent_padding_size = latent_paddings[k] * latent_window_size
+
+            indices = torch.arange(0, sum([1, latent_padding_size, latent_window_size, *history_sizes]))
+            (
+                indices_prefix,
+                indices_padding,
+                indices_latents,
+                indices_postfix,
+                indices_latents_history_2x,
+                indices_latents_history_4x,
+            ) = indices.split([1, latent_padding_size, latent_window_size, *history_sizes], dim=0)
+            # Inverted anti-drifting sampling: Figure 2(c) in the paper
+            indices_clean_latents = torch.cat([indices_prefix, indices_postfix], dim=0)
+
+            latents_prefix = image_latents
+            latents_postfix, latents_history_2x, latents_history_4x = history_latents[
+                :, :, : sum(history_sizes)
+            ].split(history_sizes, dim=2)
+            if last_image is not None and is_first_section:
+                latents_postfix = last_image_latents
+            latents_clean = torch.cat([latents_prefix, latents_postfix], dim=2)
+
+            latents = self.prepare_latents(
+                batch_size,
+                num_channels_latents,
+                height,
+                width,
+                window_num_frames,
+                dtype=torch.float32,
+                device=device,
+                generator=generator,
+                latents=None,
+            )
+
+            sigmas = np.linspace(1.0, 0.0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas
+            image_seq_len = (
+                latents.shape[2] * latents.shape[3] * latents.shape[4] / self.transformer.config.patch_size**2
+            )
+            exp_max = 7.0
+            mu = calculate_shift(
+                image_seq_len,
+                self.scheduler.config.get("base_image_seq_len", 256),
+                self.scheduler.config.get("max_image_seq_len", 4096),
+                self.scheduler.config.get("base_shift", 0.5),
+                self.scheduler.config.get("max_shift", 1.15),
+            )
+            mu = min(mu, math.log(exp_max))
+            timesteps, num_inference_steps = retrieve_timesteps(
+                self.scheduler, num_inference_steps, device, sigmas=sigmas, mu=mu
+            )
+            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
+                    timestep = t.expand(latents.shape[0])
+
+                    noise_pred = self.transformer(
+                        hidden_states=latents.to(transformer_dtype),
+                        timestep=timestep,
+                        encoder_hidden_states=prompt_embeds,
+                        encoder_attention_mask=prompt_attention_mask,
+                        pooled_projections=pooled_prompt_embeds,
+                        image_embeds=image_embeds,
+                        indices_latents=indices_latents,
+                        guidance=guidance,
+                        latents_clean=latents_clean.to(transformer_dtype),
+                        indices_latents_clean=indices_clean_latents,
+                        latents_history_2x=latents_history_2x.to(transformer_dtype),
+                        indices_latents_history_2x=indices_latents_history_2x,
+                        latents_history_4x=latents_history_4x.to(transformer_dtype),
+                        indices_latents_history_4x=indices_latents_history_4x,
+                        attention_kwargs=attention_kwargs,
+                        return_dict=False,
+                    )[0]
+
+                    if do_true_cfg:
+                        neg_noise_pred = self.transformer(
+                            hidden_states=latents.to(transformer_dtype),
+                            timestep=timestep,
+                            encoder_hidden_states=negative_prompt_embeds,
+                            encoder_attention_mask=negative_prompt_attention_mask,
+                            pooled_projections=negative_pooled_prompt_embeds,
+                            image_embeds=image_embeds,
+                            indices_latents=indices_latents,
+                            guidance=guidance,
+                            latents_clean=latents_clean.to(transformer_dtype),
+                            indices_latents_clean=indices_clean_latents,
+                            latents_history_2x=latents_history_2x.to(transformer_dtype),
+                            indices_latents_history_2x=indices_latents_history_2x,
+                            latents_history_4x=latents_history_4x.to(transformer_dtype),
+                            indices_latents_history_4x=indices_latents_history_4x,
+                            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
+                    latents = self.scheduler.step(noise_pred.float(), t, latents, return_dict=False)[0]
+
+                    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()
+
+                if is_last_section:
+                    latents = torch.cat([image_latents, latents], dim=2)
+
+                total_generated_latent_frames += latents.shape[2]
+                history_latents = torch.cat([latents, history_latents], dim=2)
+
+                real_history_latents = history_latents[:, :, :total_generated_latent_frames]
+
+                if history_video is None:
+                    if not output_type == "latent":
+                        current_latents = real_history_latents.to(vae_dtype) / self.vae.config.scaling_factor
+                        history_video = self.vae.decode(current_latents, return_dict=False)[0]
+                    else:
+                        history_video = [real_history_latents]
+                else:
+                    if not output_type == "latent":
+                        section_latent_frames = (
+                            (latent_window_size * 2 + 1) if is_last_section else (latent_window_size * 2)
+                        )
+                        overlapped_frames = (latent_window_size - 1) * self.vae_scale_factor_temporal + 1
+                        current_latents = (
+                            real_history_latents[:, :, :section_latent_frames].to(vae_dtype)
+                            / self.vae.config.scaling_factor
+                        )
+                        current_video = self.vae.decode(current_latents, return_dict=False)[0]
+                        history_video = self._soft_append(current_video, history_video, overlapped_frames)
+                    else:
+                        history_video.append(real_history_latents)
+
+        self._current_timestep = None
+
+        if not output_type == "latent":
+            generated_frames = history_video.size(2)
+            generated_frames = (
+                generated_frames - 1
+            ) // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1
+            history_video = history_video[:, :, :generated_frames]
+            video = self.video_processor.postprocess_video(history_video, output_type=output_type)
+        else:
+            video = history_video
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (video,)
+
+        return HunyuanVideoFramepackPipelineOutput(frames=video)
+
+    def _soft_append(self, history: torch.Tensor, current: torch.Tensor, overlap: int = 0):
+        if overlap <= 0:
+            return torch.cat([history, current], dim=2)
+
+        assert history.shape[2] >= overlap, f"Current length ({history.shape[2]}) must be >= overlap ({overlap})"
+        assert current.shape[2] >= overlap, f"History length ({current.shape[2]}) must be >= overlap ({overlap})"
+
+        weights = torch.linspace(1, 0, overlap, dtype=history.dtype, device=history.device).view(1, 1, -1, 1, 1)
+        blended = weights * history[:, :, -overlap:] + (1 - weights) * current[:, :, :overlap]
+        output = torch.cat([history[:, :, :-overlap], blended, current[:, :, overlap:]], dim=2)
+
+        return output.to(history)

src/diffusers/pipelines/hunyuan_video/pipeline_output.py

 from dataclasses import dataclass
+from typing import List, Union
 
+import numpy as np
+import PIL.Image
 import torch
 
 from diffusers.utils import BaseOutput
@@ -18,3 +21,19 @@ class HunyuanVideoPipelineOutput(BaseOutput):
     """
 
     frames: torch.Tensor
+
+
+@dataclass
+class HunyuanVideoFramepackPipelineOutput(BaseOutput):
+    r"""
+    Output class for HunyuanVideo pipelines.
+
+    Args:
+        frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+            List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing
+            denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+            `(batch_size, num_frames, channels, height, width)`. Or, a list of torch tensors where each tensor
+            corresponds to a latent that decodes to multiple frames.
+    """
+
+    frames: Union[torch.Tensor, np.ndarray, List[List[PIL.Image.Image]], List[torch.Tensor]]

src/diffusers/utils/dummy_pt_objects.py

@@ -565,6 +565,21 @@ class HunyuanDiT2DMultiControlNetModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class HunyuanVideoFramepackTransformer3DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class HunyuanVideoTransformer3DModel(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -692,6 +692,21 @@ class HunyuanSkyreelsImageToVideoPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class HunyuanVideoFramepackPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class HunyuanVideoImageToVideoPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/hunyuan_video/test_hunyuan_video_framepack.py

+# Copyright 2025 The HuggingFace Team.
+#
+# 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.
+
+import inspect
+import unittest
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers import (
+    CLIPTextConfig,
+    CLIPTextModel,
+    CLIPTokenizer,
+    LlamaConfig,
+    LlamaModel,
+    LlamaTokenizer,
+    SiglipImageProcessor,
+    SiglipVisionModel,
+)
+
+from diffusers import (
+    AutoencoderKLHunyuanVideo,
+    FasterCacheConfig,
+    FlowMatchEulerDiscreteScheduler,
+    HunyuanVideoFramepackPipeline,
+    HunyuanVideoFramepackTransformer3DModel,
+)
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    torch_device,
+)
+
+from ..test_pipelines_common import (
+    FasterCacheTesterMixin,
+    PipelineTesterMixin,
+    PyramidAttentionBroadcastTesterMixin,
+    to_np,
+)
+
+
+enable_full_determinism()
+
+
+class HunyuanVideoFramepackPipelineFastTests(
+    PipelineTesterMixin, PyramidAttentionBroadcastTesterMixin, FasterCacheTesterMixin, unittest.TestCase
+):
+    pipeline_class = HunyuanVideoFramepackPipeline
+    params = frozenset(
+        ["image", "prompt", "height", "width", "guidance_scale", "prompt_embeds", "pooled_prompt_embeds"]
+    )
+    batch_params = frozenset(["image", "prompt"])
+    required_optional_params = frozenset(
+        [
+            "num_inference_steps",
+            "generator",
+            "return_dict",
+            "callback_on_step_end",
+            "callback_on_step_end_tensor_inputs",
+        ]
+    )
+
+    supports_dduf = False
+    # there is no xformers processor for Flux
+    test_xformers_attention = False
+    test_layerwise_casting = True
+    test_group_offloading = True
+
+    faster_cache_config = FasterCacheConfig(
+        spatial_attention_block_skip_range=2,
+        spatial_attention_timestep_skip_range=(-1, 901),
+        unconditional_batch_skip_range=2,
+        attention_weight_callback=lambda _: 0.5,
+        is_guidance_distilled=True,
+    )
+
+    def get_dummy_components(self, num_layers: int = 1, num_single_layers: int = 1):
+        torch.manual_seed(0)
+        transformer = HunyuanVideoFramepackTransformer3DModel(
+            in_channels=4,
+            out_channels=4,
+            num_attention_heads=2,
+            attention_head_dim=10,
+            num_layers=num_layers,
+            num_single_layers=num_single_layers,
+            num_refiner_layers=1,
+            patch_size=2,
+            patch_size_t=1,
+            guidance_embeds=True,
+            text_embed_dim=16,
+            pooled_projection_dim=8,
+            rope_axes_dim=(2, 4, 4),
+            image_condition_type=None,
+            has_image_proj=True,
+            image_proj_dim=32,
+            has_clean_x_embedder=True,
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderKLHunyuanVideo(
+            in_channels=3,
+            out_channels=3,
+            latent_channels=4,
+            down_block_types=(
+                "HunyuanVideoDownBlock3D",
+                "HunyuanVideoDownBlock3D",
+                "HunyuanVideoDownBlock3D",
+                "HunyuanVideoDownBlock3D",
+            ),
+            up_block_types=(
+                "HunyuanVideoUpBlock3D",
+                "HunyuanVideoUpBlock3D",
+                "HunyuanVideoUpBlock3D",
+                "HunyuanVideoUpBlock3D",
+            ),
+            block_out_channels=(8, 8, 8, 8),
+            layers_per_block=1,
+            act_fn="silu",
+            norm_num_groups=4,
+            scaling_factor=0.476986,
+            spatial_compression_ratio=8,
+            temporal_compression_ratio=4,
+            mid_block_add_attention=True,
+        )
+
+        torch.manual_seed(0)
+        scheduler = FlowMatchEulerDiscreteScheduler(shift=7.0)
+
+        llama_text_encoder_config = LlamaConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=16,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=2,
+            pad_token_id=1,
+            vocab_size=1000,
+            hidden_act="gelu",
+            projection_dim=32,
+        )
+        clip_text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=8,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=2,
+            pad_token_id=1,
+            vocab_size=1000,
+            hidden_act="gelu",
+            projection_dim=32,
+        )
+
+        torch.manual_seed(0)
+        text_encoder = LlamaModel(llama_text_encoder_config)
+        tokenizer = LlamaTokenizer.from_pretrained("finetrainers/dummy-hunyaunvideo", subfolder="tokenizer")
+
+        torch.manual_seed(0)
+        text_encoder_2 = CLIPTextModel(clip_text_encoder_config)
+        tokenizer_2 = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+
+        feature_extractor = SiglipImageProcessor.from_pretrained(
+            "hf-internal-testing/tiny-random-SiglipVisionModel", size={"height": 30, "width": 30}
+        )
+        image_encoder = SiglipVisionModel.from_pretrained("hf-internal-testing/tiny-random-SiglipVisionModel")
+
+        components = {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "text_encoder_2": text_encoder_2,
+            "tokenizer": tokenizer,
+            "tokenizer_2": tokenizer_2,
+            "feature_extractor": feature_extractor,
+            "image_encoder": image_encoder,
+        }
+        return components
+
+    def get_dummy_inputs(self, device, seed=0):
+        if str(device).startswith("mps"):
+            generator = torch.manual_seed(seed)
+        else:
+            generator = torch.Generator(device=device).manual_seed(seed)
+
+        image_height = 32
+        image_width = 32
+        image = Image.new("RGB", (image_width, image_height))
+        inputs = {
+            "image": image,
+            "prompt": "dance monkey",
+            "prompt_template": {
+                "template": "{}",
+                "crop_start": 0,
+            },
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 4.5,
+            "height": image_height,
+            "width": image_width,
+            "num_frames": 9,
+            "latent_window_size": 3,
+            "max_sequence_length": 256,
+            "output_type": "pt",
+        }
+        return inputs
+
+    def test_inference(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        video = pipe(**inputs).frames
+        generated_video = video[0]
+
+        self.assertEqual(generated_video.shape, (13, 3, 32, 32))
+        expected_video = torch.randn(13, 3, 32, 32)
+        max_diff = np.abs(generated_video - expected_video).max()
+        self.assertLessEqual(max_diff, 1e10)
+
+    def test_callback_inputs(self):
+        sig = inspect.signature(self.pipeline_class.__call__)
+        has_callback_tensor_inputs = "callback_on_step_end_tensor_inputs" in sig.parameters
+        has_callback_step_end = "callback_on_step_end" in sig.parameters
+
+        if not (has_callback_tensor_inputs and has_callback_step_end):
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe = pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+        self.assertTrue(
+            hasattr(pipe, "_callback_tensor_inputs"),
+            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        )
+
+        def callback_inputs_subset(pipe, i, t, callback_kwargs):
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        def callback_inputs_all(pipe, i, t, callback_kwargs):
+            for tensor_name in pipe._callback_tensor_inputs:
+                assert tensor_name in callback_kwargs
+
+            # iterate over callback args
+            for tensor_name, tensor_value in callback_kwargs.items():
+                # check that we're only passing in allowed tensor inputs
+                assert tensor_name in pipe._callback_tensor_inputs
+
+            return callback_kwargs
+
+        inputs = self.get_dummy_inputs(torch_device)
+
+        # Test passing in a subset
+        inputs["callback_on_step_end"] = callback_inputs_subset
+        inputs["callback_on_step_end_tensor_inputs"] = ["latents"]
+        output = pipe(**inputs)[0]
+
+        # Test passing in a everything
+        inputs["callback_on_step_end"] = callback_inputs_all
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+
+        def callback_inputs_change_tensor(pipe, i, t, callback_kwargs):
+            is_last = i == (pipe.num_timesteps - 1)
+            if is_last:
+                callback_kwargs["latents"] = torch.zeros_like(callback_kwargs["latents"])
+            return callback_kwargs
+
+        inputs["callback_on_step_end"] = callback_inputs_change_tensor
+        inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
+        output = pipe(**inputs)[0]
+        assert output.abs().sum() < 1e10
+
+    def test_attention_slicing_forward_pass(
+        self, test_max_difference=True, test_mean_pixel_difference=True, expected_max_diff=1e-3
+    ):
+        if not self.test_attention_slicing:
+            return
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        generator_device = "cpu"
+        inputs = self.get_dummy_inputs(generator_device)
+        output_without_slicing = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=1)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing1 = pipe(**inputs)[0]
+
+        pipe.enable_attention_slicing(slice_size=2)
+        inputs = self.get_dummy_inputs(generator_device)
+        output_with_slicing2 = pipe(**inputs)[0]
+
+        if test_max_difference:
+            max_diff1 = np.abs(to_np(output_with_slicing1) - to_np(output_without_slicing)).max()
+            max_diff2 = np.abs(to_np(output_with_slicing2) - to_np(output_without_slicing)).max()
+            self.assertLess(
+                max(max_diff1, max_diff2),
+                expected_max_diff,
+                "Attention slicing should not affect the inference results",
+            )
+
+    def test_vae_tiling(self, expected_diff_max: float = 0.2):
+        # Seems to require higher tolerance than the other tests
+        expected_diff_max = 0.6
+        generator_device = "cpu"
+        components = self.get_dummy_components()
+
+        pipe = self.pipeline_class(**components)
+        pipe.to("cpu")
+        pipe.set_progress_bar_config(disable=None)
+
+        # Without tiling
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_without_tiling = pipe(**inputs)[0]
+
+        # With tiling
+        pipe.vae.enable_tiling(
+            tile_sample_min_height=96,
+            tile_sample_min_width=96,
+            tile_sample_stride_height=64,
+            tile_sample_stride_width=64,
+        )
+        inputs = self.get_dummy_inputs(generator_device)
+        inputs["height"] = inputs["width"] = 128
+        output_with_tiling = pipe(**inputs)[0]
+
+        self.assertLess(
+            (to_np(output_without_tiling) - to_np(output_with_tiling)).max(),
+            expected_diff_max,
+            "VAE tiling should not affect the inference results",
+        )
+
+    # TODO(aryan): Create a dummy gemma model with smol vocab size
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_consistent(self):
+        pass
+
+    @unittest.skip(
+        "A very small vocab size is used for fast tests. So, any kind of prompt other than the empty default used in other tests will lead to a embedding lookup error. This test uses a long prompt that causes the error."
+    )
+    def test_inference_batch_single_identical(self):
+        pass