LongCat Video

LongCat-Video is a comprehensive video generation model by meituan-longcat developed on GitHub. This AI tool is designed to execute multiple tasks within one video generation framework, including converting text to video, translating images into video sequences, and generating continuation in videos. One key feature of LongCat-Video is in its ability to efficiently create long videos without reduction in quality or noticeable color drifting. It follows a coarse-to-fine generation strategy along both temporal and spatial axes to enhance efficiency, especially at high resolutions. The model was trained through Group Relative Policy Optimization, which incorporates multi-reward Real-life High Fidelity (RLHF) and ensures competitive performance across multiple metrics when compared to leading open-source and commercial video generation models. Moreover, the AI model has launched an expressive audio-driven character animation feature, known as LongCat-Video-Avatar, which can natively handle tasks such as Audio-Text-to-Video conversion, Audio-Text-Image-to-Video generation, and Video Continuation. It offers seamless compatibility for both single-stream and multi-stream audio inputs. All technical reports, inference code, model weights, and project pages related to LongCat-Video are openly available on GitHub.