Teaching diffusion models to keep the camera rolling

What it does SkyReels-V2 is a suite of open-weights video generation models built around an autoregressive diffusion-forcing architecture. It produces long-form video by autoregressively extending clips with overlapping frame history, handling both text-to-video and image-to-video tasks at 540P and 720P. The repository bundles inference code, a video captioning model named SkyCaptioner-V1, and a prompt enhancer.

The interesting bit The trick is treating generation like a relay race: each new chunk conditions on prior frames so the model can, in theory, keep rolling indefinitely. The authors claim this is the first open-source autoregressive diffusion-forcing system to hit state-of-the-art results among publicly available models.

Key highlights

• Released 1.3B and 14B parameter checkpoints; 5B variants and a “Camera Director” mode are still on the roadmap.
• Supports synchronous and asynchronous inference for long-form generation, plus start/end frame control and video extension.
• Integrates with the diffusers library and supports both single-GPU and multi-GPU inference.
• Ships with auxiliary tooling: SkyCaptioner-V1 for video captioning and a built-in prompt enhancer.

Caveats

• Several promised checkpoints—including the 5B series, Camera Director models, and Step & Guidance Distill variants—remain marked “Coming Soon.”
• The repository is strictly an inference release; training code and detailed architectural documentation are not included in the README.

Verdict Worth exploring if you need open weights for long-form video synthesis and can live with the 14B or 1.3B tiers. Look elsewhere if you want training frameworks or a fully released model catalog.