Meituan Releases LongCat-Video as Open-Source Model
Meituan has released its LongCat-Video model as open-source. The model can generate video from text, images, and existing video clips.
Chinese technology company Meituan has launched its LongCat-Video model under an open-source MIT license, allowing for widespread use and modification.
The model is capable of generating video content from various inputs, including text prompts, images, and by continuing existing video sequences. A key feature highlighted is its ability to produce videos lasting several minutes, a notable challenge for many prior open-source models.
LongCat-Video operates on a dense Transformer architecture with 13.6 billion parameters. It supports text-to-video (T2V), image-to-video (I2V), and video continuation tasks within a unified framework. Its video continuation capability is particularly noted for avoiding common issues like color drift and motion inconsistency.
The model's efficiency is attributed to its Block Sparse Attention mechanism, which enables the generation of longer videos without an exponential increase in computational cost. Meituan's decision to open-source the model aims to foster community development and accelerate innovation in video generation technology.
The model is available on GitHub and Hugging Face, with its technical report also published for developers.