Add SkyReels V2: Infinite-Length Film Generative Model

[tolgacangoz]

Thanks for the opportunity to fix #11374!

Original Work

Original repo: https://github.com/SkyworkAI/SkyReels-V2
Paper: https://huggingface.co/papers/2504.13074

SkyReels V2's main contributions are summarized as follow:
• Comprehensive video captioner that understand the shot language while capturing the general description of the video, which dramatically improve the prompt adherence.
• Motion-specific preference optimization enhances motion dynamics with a semi-automatic data collection pipeline.
• Effective Diffusion-forcing adaptation enables the generation of ultra-long videos and story generation capabilities, providing a robust framework for extending temporal coherence and narrative depth.
• SkyCaptioner-V1 and SkyReels-V2 series models including diffusion-forcing, text2video, image2video, camera director and elements2video models with various sizes (1.3B, 5B, 14B) are open-sourced.

TODOs:
✅ SkyReelsV2Transformer3DModel: 90% WanTransformer3DModel
✅ SkyReelsV2DiffusionForcingPipeline
✅ SkyReelsV2DiffusionForcingImageToVideoPipeline: Includes FLF2V.
✅ SkyReelsV2DiffusionForcingVideoToVideoPipeline: Extends a given video.
✅ SkyReelsV2Pipeline
✅ SkyReelsV2ImageToVideoPipeline: Includes FLF2V.
✅ scripts/convert_skyreelsv2_to_diffusers.py

• tolgacangoz/SkyReels-V2-Diffusers

✅ Did you make sure to update the documentation with your changes? Did you write any new necessary tests?: We will construct these during review.

T2V with Diffusion Forcing (OLD)

Skywork/SkyReels-V2-DF-1.3B-540P
seed 0 and num_frames 97
Original repo	diffusers integration
original_0_short.mp4	diffusers_0_short.mp4

seed 37 and num_frames 97
Original repo	diffusers integration
original_37_short.mp4	diffusers_37_short.mp4

seed 0 and num_frames 257
Original repo	diffusers integration
original_0_long.mp4	diffusers_0_long.mp4

seed 37 and num_frames 257
Original repo	diffusers integration
original_37_long.mp4	diffusers_37_long.mp4

!pip install git+https://github.com/tolgacangoz/diffusers.git@skyreels-v2 ftfy -q
import torch
from diffusers import AutoencoderKLWan, SkyReelsV2DiffusionForcingPipeline
from diffusers.utils import export_to_video

vae = AutoencoderKLWan.from_pretrained(
			"tolgacangoz/SkyReels-V2-DF-1.3B-540P-Diffusers",
			subfolder="vae",
			torch_dtype=torch.float32)
pipe = SkyReelsV2DiffusionForcingPipeline.from_pretrained(
			"tolgacangoz/SkyReels-V2-DF-1.3B-540P-Diffusers",
			vae=vae,
			torch_dtype=torch.bfloat16)
pipe = pipe.to("cuda")
pipe.transformer.set_ar_attention(causal_block_size=5)

prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window."

output = pipe(
    prompt=prompt,
    num_inference_steps=30,
    height=544,
    width=960,
    num_frames=97,
    ar_step=5,  # Controls asynchronous inference (0 for synchronous mode)
    generator=torch.Generator(device="cpu").manual_seed(0),
    overlap_history=None,  # Number of frames to overlap for smooth transitions in long videos; 17 for long
    addnoise_condition=20,  # Improves consistency in long video generation
).frames[0]
export_to_video(output, "T2V.mp4", fps=24, quality=8)

"""
You can set `ar_step=5` to enable asynchronous inference. When asynchronous inference,
`causal_block_size=5` is recommended while it is not supposed to be set for
synchronous generation. Asynchronous inference will take more steps to diffuse the
whole sequence which means it will be SLOWER than synchronous mode. In our
experiments, asynchronous inference may improve the instruction following and visual consistent performance.
"""

I2V with Diffusion Forcing (OLD)

prompt="A penguin dances."	diffusers integration
	i2v-short.mp4

#!pip uninstall diffusers -yq
#!pip install git+https://github.com/tolgacangoz/diffusers.git@skyreels-v2 ftfy -q
import torch
from diffusers import AutoencoderKLWan, SkyReelsV2DiffusionForcingImageToVideoPipeline
from diffusers.utils import export_to_video, load_image

vae = AutoencoderKLWan.from_pretrained(
			"tolgacangoz/SkyReels-V2-DF-1.3B-540P-Diffusers",
			subfolder="vae",
			torch_dtype=torch.float32)
pipe = SkyReelsV2DiffusionForcingImageToVideoPipeline.from_pretrained(
			"tolgacangoz/SkyReels-V2-DF-1.3B-540P-Diffusers",
			vae=vae,
			torch_dtype=torch.bfloat16)
pipe = pipe.to("cuda")
#pipe.transformer.set_ar_attention(causal_block_size=5)

image = load_image("Penguin from https://huggingface.co/tasks/image-to-video")
prompt = "A penguin dances."

output = pipe(
    image=image,
    prompt=prompt,
    num_inference_steps=50,
    height=544,
    width=960,
    num_frames=97,
    #ar_step=5,  # Controls asynchronous inference (0 for synchronous mode)
    generator=torch.Generator(device="cpu").manual_seed(0),
    overlap_history=None,  # Number of frames to overlap for smooth transitions in long videos; 17 for long
    addnoise_condition=20,  # Improves consistency in long video generation
).frames[0]
export_to_video(output, "I2V.mp4", fps=24, quality=8)

"""
When I set `ar_step=5` and `causal_block_size=5`, then the results seem really bad.
"""

FLF2V with Diffusion Forcing (OLD)

Now, Houston, we have a problem.
I have been unable to produce good results with this task. I tried many hyperparameter combinations with the original code.
The first frame's latent (torch.Size([1, 16, 1, 68, 120])) is overwritten onto the first of 25 frame latents of latents (torch.Size([1, 16, 25, 68, 120])). Then, the last frame's latent is concatenated, thus latents is torch.Size([1, 16, 26, 68, 120]). After the denoising process, the length of the last frame latent is discarded at the end and then decoded by the VAE. I tried not concatenating the last frame but overwriting onto the latest frame of latents and not discarding the latest frame latent at the end, but still got bad results. Here are some results:

First Frame	Last Frame
0.mp4	1.mp4
2.mp4	3.mp4
4.mp4	5.mp4
6.mp4	7.mp4

#!pip uninstall diffusers -yq
#!pip install git+https://github.com/tolgacangoz/diffusers.git@skyreels-v2 ftfy -q
import torch
from diffusers import AutoencoderKLWan, SkyReelsV2DiffusionForcingImageToVideoPipeline
from diffusers.utils import export_to_video, load_image

vae = AutoencoderKLWan.from_pretrained(
			"tolgacangoz/SkyReels-V2-DF-1.3B-540P-Diffusers",
			subfolder="vae",
			torch_dtype=torch.float32)
pipe = SkyReelsV2DiffusionForcingImageToVideoPipeline.from_pretrained(
			"tolgacangoz/SkyReels-V2-DF-1.3B-540P-Diffusers",
			vae=vae,
			torch_dtype=torch.bfloat16)
pipe = pipe.to("cuda")
#pipe.transformer.set_ar_attention(causal_block_size=5)

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."
negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
first_frame = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png")
last_frame = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png")

output = pipe(
    image=first_frame,
    last_image=last_frame,
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=50,
    height=544,
    width=960,
    num_frames=97,
    #ar_step=5,  # Controls asynchronous inference (0 for synchronous mode)
    generator=torch.Generator(device="cpu").manual_seed(0),
    overlap_history=None,  # Number of frames to overlap for smooth transitions in long videos; 17 for long
    addnoise_condition=20,  # Improves consistency in long video generation
).frames[0]
export_to_video(output, "FLF2V.mp4", fps=24, quality=8)

V2V with Diffusion Forcing (OLD)

This pipeline extends a given video.

Input Video	diffusers integration
video1.mp4	v2v.mp4

#!pip uninstall diffusers -yq
#!pip install git+https://github.com/tolgacangoz/diffusers.git@skyreels-v2 ftfy -q
import torch
from diffusers import AutoencoderKLWan, SkyReelsV2DiffusionForcingVideoToVideoPipeline
from diffusers.utils import export_to_video, load_video

vae = AutoencoderKLWan.from_pretrained(
			"tolgacangoz/SkyReels-V2-DF-1.3B-540P-Diffusers",
			subfolder="vae",
			torch_dtype=torch.float32)
pipe = SkyReelsV2DiffusionForcingVideoToVideoPipeline.from_pretrained(
			"tolgacangoz/SkyReels-V2-DF-1.3B-540P-Diffusers",
			vae=vae,
			torch_dtype=torch.bfloat16)
pipe = pipe.to("cuda")
#pipe.transformer.set_ar_attention(causal_block_size=5)

prompt = "CG animation style, a small blue bird flaps 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 continuing flight and the vastness of the sky from a close-up, low-angle perspective."
negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"
video = load_video("Input video.mp4")

output = pipe(
    video=video,
    prompt=prompt,
    num_inference_steps=50,
    height=544,
    width=960,
    num_frames=120,
    base_num_frames=97,
    ar_step=0,  # Controls asynchronous inference (0 for synchronous mode)
    generator=torch.Generator(device="cpu").manual_seed(0),
    overlap_history=17,  # Number of frames to overlap for smooth transitions in long videos
    addnoise_condition=20,  # Improves consistency in long video generation
).frames[0]
export_to_video(output, "V2V.mp4", fps=24, quality=8)

Firstly, I want to congratulate you on this great work, and thanks for open-sourcing it, SkyReels Team! This PR proposes an integration of your model.

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag members/contributors who may be interested in your PR.

@yiyixuxu @a-r-r-o-w @linoytsaban @yjp999 @Howe2018 @RoseRollZhu @pftq @Langdx @guibinchen @qiudi0127 @nitinmukesh @tin2tin @ukaprch @okaris

[ukaprch]

It's about time. Thanks.

[tolgacangoz]

Hi @yiyixuxu @a-r-r-o-w

Mid-PR questions:

1. The issue was labelled as "contributions-welcome" but not as "community-examples". Also, the number of stars in this model surpassed that of SkyReels-V1. Thus, I located these pipelines in src/diffusers/pipelines/skyreels_v2/. Should I move these pipelines to examples/? Should I also split this PR for each pipeline (group)?

2. Just like SkyReels-V1 was based on HunyuanVideo, SkyReels-V2 is based on Wan, but some differences exist. I thought of moving the differences to the parent abstraction, i.e., pipeline code, so that we can use WanTransformer3DModel for both, but it didn't seem appropriate enough to me at first. But then, if we introduce Diffusion Forcing and AutoRegressive properties into WanTransformer3DModel (as native as possible, not with the exact diff below), it seems possible to me. You can examine the current diff between transformer_wan.py and transformer_skyreels_v2.py: https://www.diffchecker.com/U72HJ6ox/ WDYT?

3. Since SkyReels-V2 is not a completely new architecture, should I move its pipelines into src/diffusers/pipelines/wan/ similar to HunyuanSkyreelsImageToVideoPipeline, if SkyReels-V2 is seen as an official pipeline?

4. I am removing TeaCache-related code because it is planned for a modular extension, right? If this PR is required to move to examples/, then no need to remove, I think.

5. I came across this:
   

   diffusers/src/diffusers/models/embeddings.py

   Line 1153 in 01abfc8

   / (theta ** (torch.arange(0, dim, 2, dtype=freqs_dtype, device=pos.device)[: (dim // 2)] / dim))

   
   At first, [: (dim // 2)] confused me :S Isn't it redundant? dim was already confirmed even with assert dim % 2 == 0. Can I remove [: (dim // 2)] in a separate PR?

[a-r-r-o-w]

@tolgacangoz Thanks for working on this, really cool work so far!

1. I think we should add SkyReels models in core diffusers, so src/ is fine.

2 and 3. I think in this case, we should have separate implementation of SkyReelsV2 and Wan due to the autoregressive nature of the former. Adding any extra code in Wan might complicate it for readers. Will let @yiyixuxu comment on this though

4. Yeah let's remove the cache code. We'll try to write a model agnostic implementation in future once more of the cache related code is stablized after adding some of the easier methods that are not too model intrusive (such as first block cache).

5. You're right, it's redundant. Let's remove in a separate PR

[ukaprch]

FWIW, I have been successful in using the same T5 encoder for WAN 2.1 for this model just by fiddling with their pipeline:

        print('Quantize text_encoder qint8')
        class  QuantizedT5EncoderModelForCausalLM (QuantizedTransformersModel):
            auto_class = UMT5EncoderModel
            auto_class.from_config = auto_class._from_config
        text_encoder = QuantizedT5EncoderModelForCausalLM.from_pretrained(
            "./wan quantro T2V-14B-720P Diffusers/basemodel/t5encodermodel_qint8"
        ).to(dtype=dtype)
        
        pipe = Text2VideoPipeline(
            model_path=model_path,
            transformer=transformer,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            weight_dtype=dtype)

Then this: I incorporate my bitsandbytes nf4 transformer, their tokenizer and the WAN based T5 encoder:

def __init__(
    self, model_path, transformer, text_encoder, tokenizer, device: str = "cuda", weight_dtype=torch.bfloat16, use_usp=False, offload=False,
):
    self.transformer = transformer          #get_transformer(model_path, 'cpu', weight_dtype)
    vae_model_path = os.path.join(model_path, "Wan2.1_VAE.pth")
    self.vae = get_vae(vae_model_path, 'cpu', weight_dtype=torch.float32)
    if text_encoder is not None:
        self.text_encoder = text_encoder        #get_text_encoder(model_path, 'cuda', weight_dtype)
    if tokenizer is not None: 
        self.tokenizer = tokenizer

I need to add this function to the pipeline for the T5 encoder to work:

def encode(self, texts):
    ids, mask = self.tokenizer(texts, return_mask=True, add_special_tokens=True)
    ids = ids.to(self.device)
    mask = mask.to(self.device)
    context = self.text_encoder(ids, mask)
    #seq_lens = mask.gt(0).sum(dim=1).long()
    context = context.last_hidden_state * mask.unsqueeze(-1)
    return context

[tolgacangoz]

It seems appropriate to me. Only Diffusion Forcing pipelines are different for large models. How are the results with your setting?

[tolgacangoz]

Hi @yiyixuxu @a-r-r-o-w and SkyReels Team @yjp999 @pftq @Langdx @guibinchen ...

This PR will be ready for review for SkyReelsV2Transformer3DModel and SkyReelsV2DiffusionForcingPipeline soon. Other pipelines will follow quickly after initial feedback...

[tolgacangoz]

This will be my 3. pipeline contribution, yay 🥳!

[yiyixuxu]

@tolgacangoz ohh this cannot be the only change in scheduler, no?

[yiyixuxu]

ohh it's already in!

[yiyixuxu]

is the output quality match?

[tolgacangoz]

The outputs are qualitatively/visibly the same.

[yiyixuxu]

thanks!

[a-r-r-o-w]

Thanks for the amazing work here @tolgacangoz! I think the only blocker is having the weights merged into the official repos, yes?

[tolgacangoz]

Right.

[yiyixuxu]

hi @tolgacangoz

can you send PR into the official repo for the weights, I think they have created place holder for all the checkpoints, e.g.

https://huggingface.co/Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers

[tolgacangoz]

I thought they were supposed to do this by examining/verifying the conversion script, etc., since we talk about the official repository.
Sorry for the misunderstanding, working on it.

[tolgacangoz]

Or, I think they can stay as a meaning of placeholder or potential feature, because the original code was the one that I cannot produce good results with 1.3B models for FLF2V. Or, it was I who couldn't run this task properly, idk :S. Maybe it is OK with larger models. I think this PR is well-suited for its job for integration.

Edit: I opened an issue at the original repo about this. I forgot to open earlier, sry 🥲.

They say try with 14B models for FLF2V, thus this issue (?) is irrelevant from this PR, IMO.

[yiyixuxu]

@tolgacangoz
can you point me to where this conversation is so I can get some context? 😛

They say try with 14B models for FLF2V, thus this issue (?) is irrelevant from this PR, IMO.

[tolgacangoz]

If you examine the first comment of this PR, you can see that I wasn't able to produce good results for FLF2V with the DF 1.3B model by using the original code. This is their answer: SkyworkAI/SkyReels-V2#93

[yiyixuxu]

ohh sounds good, thanks for explaining!
would you be able add a code example in the README in the repo PRs?

[yiyixuxu]

thanks @tolgacangoz
merging the code PR now, will see to the weights PR too

[tolgacangoz]

Thanks for merging and for the opportunity to contribute!

I'll be monitoring the original repository for updates...

[asomoza]

thanks a lot @tolgacangoz, really awesome contribution!

[nitinmukesh]

Thank you, @tolgacangoz