On January 29,昆仑万维 (Skywork AI) officially announced the open-source release of its self-developed "versatile" video generation large model - SkyReels-V3. The release of this model marks a new stage in video generation technology, characterized by high-fidelity and multimodal capabilities. It achieves deep integration of three core functions: image-to-video (I2V), video extension, and audio-driven virtual avatars, all through a single modeling architecture.

In the field of image-to-video (I2V), SkyReels-V3 supports input of 1 to 4 reference images, accurately preserving the identity features and spatial composition of the subject. Evaluation data shows that it performs excellently in reference consistency and visual quality metrics, surpassing mainstream commercial models such as Vidu Q2 and Kling1.6.

QQ20260129-154437.png

Addressing the issues of duration and narrative in video creation, SkyReels-V3 introduces a powerful video extension function. This function not only supports simple perspective continuation but also introduces a "shot transition extension mode," incorporating professional cinematic transition techniques such as cut-in and front-back shots, enabling AI-generated videos to evolve from simple "time expansion" to "narrative expansion" with logic.

Additionally, the audio-driven virtual avatar (Talking Avatar) module achieves a high lip-sync rate through precise audio-video alignment technology, supporting minute-level long video generation and multi-character interaction, providing a more natural interactive experience for digital human live streaming and online education.

QQ20260129-154523.png

Currently, SkyReels-V3 has been officially open-sourced on GitHub, and provides free API call service for a limited time. Skywork AI stated that it will continue to focus on the "All in AGI and AIGC" strategy, promoting the prosperity of the AI video creation ecosystem through open source.

Address: GitHub: https://github.com/SkyworkAI/SkyReels-V3