wtf is longcat-video?

LongCat-Video is an AI video generation model built by Meituan's LongCat team, with 13.6 billion parameters. It can turn text descriptions into video, turn a single image into a video, and continue an existing video clip, all using one unified model rather than separate tools for each task. Its main strength is generating long videos, several minutes in length, without the color shifting or quality loss that often shows up in AI generated video over time. It produces 720p video at 30 frames per second and uses techniques like a coarse to fine generation strategy and Block Sparse Attention to keep inference fast even at higher resolutions. The model was trained using a technique called multi reward Group Relative Policy Optimization, and the team reports it performs comparably to both other open source video models and commercial alternatives. The project also includes a related model called LongCat-Video-Avatar, which generates audio driven character animation, meaning a video of a character whose lips and movements sync to a given audio track. A newer version, Avatar 1.5, improves lip synchronization accuracy, works better across stylized art styles like anime, and runs faster through a distillation technique that cuts inference down to eight steps. To use it, you clone the repository, set up a conda environment with Python 3.10, install PyTorch configured for your CUDA version along with flash-attention, and download the pretrained model weights from Hugging Face. From there you can run text to video generation on either a single GPU or multiple GPUs using the included demo scripts. The project is written in Python and released under the MIT license.