SkyReels-V2 is an infinite-length film generative model that generates video using an autoregressive diffusion-forcing architecture.
The tool addresses the challenge of generating long-form video content by employing an autoregressive diffusion-forcing approach, which allows it to generate video sequences of arbitrary length. The architecture works by iteratively predicting and refining video frames in an autoregressive manner, conditioning each generation step on previously generated content. This enables the model to maintain temporal coherence across extended sequences without being constrained by fixed output lengths.
Developers should consider this tool if they need to generate long-form video content programmatically. The project provides inference code and model weights, with variants optimized for different tasks: a diffusion-forcing model for infinite-length autoregressive video generation and an image-to-video model for synthesis from static images. The tool also supports video extension capabilities and start/end frame control, making it suitable for applications requiring flexible video generation workflows. A video captioning model is included alongside the core generation models.
The project maintains active development with regular releases of new model variants and capabilities. The team has released higher-resolution model versions and expanded the framework with complementary tools for portrait animation and controllable video generation. Technical documentation is available through a published technical report, and the models are distributed across multiple platforms including Hugging Face and ModelScope. Community engagement is supported through a Discord channel and a web-based playground for testing the models.
