Generate a short video clip from a written scene description using the text-to-video mode.
Animate a still product image into moving footage using image-to-video generation.
Extend an existing video clip by generating more footage and appending it to the end.
Specify an opening and closing frame and let the model fill in the motion between them.
Requires significant GPU memory, the 14B model needs multi-GPU setup and models must be downloaded from Hugging Face before first run.
SkyReels V2 is an open-source AI video generation system from Skywork AI. Given a text description or a starting image, it generates video footage. The standout claim is that it can produce videos of any length by generating them in continuous segments rather than as a fixed-length clip. The developers describe the underlying approach as an AutoRegressive Diffusion-Forcing architecture, which is the mechanism that allows the model to keep extending the video beyond typical limits. The system comes in several sizes. The smallest is 1.3 billion parameters, and the largest publicly released version is 14 billion parameters. Models are available at 540p and 720p resolutions. There are three main generation modes: text-to-video (generate from a written description), image-to-video (animate a still image), and diffusion forcing (the mode used for long or infinite-length generation). A separate video extension mode lets you take an existing clip and add more footage to the end of it. A start-and-end frame control mode lets you specify both the opening and closing images and have the model fill in the motion between them. Running the larger models requires substantial computing resources. The README describes multi-GPU inference options and provides guidance on memory requirements. Models are downloaded from Hugging Face or ModelScope before running the included Python scripts. The repository also includes a video captioning model called SkyCaptioner-V1 and a prompt enhancement tool to help users write better generation prompts. The project is part of a broader series. SkyReels-V1 was released earlier and focused on human-centric video generation. SkyReels-V3 is now available as a separate repository. Related projects in the same organization include portrait animation tools and a controllable generation framework for assembling specific visual elements into a scene. The full README is longer than what was shown.
← skyworkai on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.