SkyReels-V2

Billed as the first infinite-length film generative model, SkyReels-V2 from Skywork AI uses a Diffusion Forcing framework that generates video autoregressively, extending clips indefinitely instead of stopping at a fixed frame budget. Released open weights come in 1.3B and 14B sizes, with a 5B series still listed as coming soon, covering text-to-video, image-to-video, video extension, and camera control at 540p (544x960, 97 frames per segment) or 720p (720x1280, 121 frames), and the release includes the SkyCaptioner-V1 video captioning model used to build its training data. Peak VRAM runs from about 14.7 GB for the 1.3B model at 540p up to 51.2 GB for the 14B with diffusion forcing, with xDiT USP multi-GPU inference and Diffusers integration available. Installation is pip install -r requirements.txt on Python 3.10. The models ship under the custom Skywork Community License, which permits commercial use subject to conditions rather than being a standard OSI license. With around 7,300 GitHub stars it is a popular base for long-form open video work, and Skywork shipped the successor SkyReels-V3 in January 2026.