SkyReels Hunyuan T2V & I2V

[a-r-r-o-w]

Adds support for SkyReels-V1. Also fixes#10768 and overrides #10567

Repo: https://github.com/SkyworkAI/SkyReels-V1

The authors used the diffusers model implementation and created the pipeline themselves, so all credits to them! This is just a quick ctrl c+v

T2V:

importtorchimporttorch._dynamo.configfromdiffusersimportHunyuanVideoPipeline, HunyuanVideoTransformer3DModelfromdiffusers.utilsimportexport_to_videotorch._dynamo.config.inline_inbuilt_nn_modules=Truemodel_id="hunyuanvideo-community/HunyuanVideo"transformer_model_id="Skywork/SkyReels-V1-Hunyuan-T2V"transformer=HunyuanVideoTransformer3DModel.from_pretrained(
transformer_model_id, torch_dtype=torch.bfloat16
)
pipe=HunyuanVideoPipeline.from_pretrained(model_id, transformer=transformer, torch_dtype=torch.float16)
pipe.vae.enable_tiling()
pipe.to("cuda")
output=pipe(
prompt="A cat walks on the grass, realistic",
negative_prompt="Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion",
height=544,
width=960,
num_frames=97,
guidance_scale=1.0,
true_cfg_scale=6.0,
num_inference_steps=30,
).frames[0]
export_to_video(output, "output.mp4", fps=15)

I2V:

importtorchfromdiffusersimportHunyuanSkyreelsImageToVideoPipeline, HunyuanVideoTransformer3DModelfromdiffusers.utilsimportload_image, export_to_videomodel_id="hunyuanvideo-community/HunyuanVideo"transformer_model_id="Skywork/SkyReels-V1-Hunyuan-I2V"transformer=HunyuanVideoTransformer3DModel.from_pretrained(
transformer_model_id, torch_dtype=torch.bfloat16
)
pipe=HunyuanSkyreelsImageToVideoPipeline.from_pretrained(
model_id, transformer=transformer, torch_dtype=torch.float16
)
pipe.vae.enable_tiling()
pipe.to("cuda")
prompt="An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot."negative_prompt="Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion"image=load_image(
"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
)
output=pipe(
image=image,
prompt=prompt,
negative_prompt=negative_prompt,
num_inference_steps=30
).frames[0]
export_to_video(output, "output_i2v.mp4", fps=15)

T2V	I2V
output.mp4	output_i2v.mp4

cc @asomoza@Langdx@Howe2018

[a-r-r-o-w]

pooled_prompt_embeds can never be None otherwise we'll error out in transformer. Change not required for integration but we should do it anyway (perhaps in separate PR if not here)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[yiyixuxu]

looks good! thank you!

[tin2tin]

Pretty heavy models to run locally. I hope will be some hints on how to run it on consumer hardware?

[nitinmukesh]

Pretty heavy models to run locally. I hope will be some hints on how to run it on consumer hardware?

int4 didn't run on 24GB?

[tin2tin]

I had to bump torch to get the dynamo working. After that I got int4 working. Thank you @nitinmukesh for the int4 version of the transformer.

Is support for Skyworks' A1 (motion transfer) considered?

[a-r-r-o-w]

Oh... the dynamo part is not required 🫠 I was just benchmarking torch compile with the model and forgot to remove when copy-pasting the minimal snippet.

Regarding Skyreels A1, could you open an issue for tracking? Currently, we haven't planned for it due to some other priorities but any community contribution from looking at an open issue would be helpful from the interested folks!

[a-r-r-o-w]

Also, ideally with group offloading (and possibly combining layerwise upcasting), we have the lowest VRAM requirements (without much generation time overhead) I believe. Currently, we're limited by CPU RAM, but we'll try to improve on that asap :)

[tin2tin]

Lol, didn't understand the last part. No worries. In general, if a new model is VRAM intensive and the Diffusers team has knowledge of ways to improve that, please include some hints in the model card.