SkyReels-V2 is a groundbreaking open-source video generative model that employs an AutoRegressive Diffusion-Forcing architecture to achieve state-of-the-art performance in video generation. It is the first publicly available model capable of generating infinite-length videos while maintaining high visual quality, motion dynamics, and prompt adherence. The project includes model weights, inference code, and a video captioning model (SkyCaptioner-V1) for detailed annotations.
Key features include:
- Infinite-length video generation: Utilizing Diffusion Forcing framework for seamless long-form synthesis.
- High-resolution outputs: Supports 540P and 720P resolutions.
- Multiple applications: Text-to-Video (T2V), Image-to-Video (I2V), and Camera Director functionality.
- Optimized inference: Includes options for synchronous and asynchronous modes, multi-GPU acceleration, and VRAM optimization.
SkyReels-V2 has been rigorously evaluated against leading models, demonstrating superior performance in instruction adherence, consistency, and visual quality. The project also provides comprehensive documentation, including technical reports, inference scripts, and integration with platforms like Hugging Face and ModelScope.