SkyReels

SkyReels-V2 is the downloadable video model family for developers and researchers who want to generate footage on their own GPUs. It includes Python inference code and model weights, distinct from the SkyReels web playground. Generation runs on your hardware.

Dedicated text-to-video models create clips from written descriptions, while image-to-video models use an image as the opening frame. The Diffusion Forcing models support both inputs and continue generation using frames from earlier segments. They can also extend an existing video or accept starting and ending images to control a sequence's endpoints.

This continuation approach makes longer sequences possible without limiting each result to one short clip. Documented examples cover videos up to 60 seconds. Diffusion Forcing offers synchronous and asynchronous generation; asynchronous processing takes longer and may improve prompt adherence and visual consistency.

Inference supports a single GPU or multiple GPUs through xDiT USP. Hugging Face Diffusers integrations cover text and image generation, as well as Diffusion Forcing video extension. Downloadable weights are available through Hugging Face and ModelScope, with 540p and 720p variants.

GPU memory needs are substantial. Reported peak use for 540p generation is about 14.7 GB for the smaller model and roughly 43 to 51 GB for the larger models, depending on the task. CPU offloading reduces VRAM use, while processing fewer frames at once can trade some visual quality for lower memory demand. The repository uses a custom license.