On October 27, Meituan's LongCat team made an official announcement, introducing the LongCat-Video model designed for video generation. Built upon the Diffusion Transformer (DiT) framework, this model categorizes tasks based on the "number of conditional frames". It inherently supports three fundamental tasks: generating videos from text, generating videos from images, and extending existing videos. These tasks collectively form a comprehensive task cycle.

In the realm of text-to-video generation, the model is capable of producing high-definition 720p videos at a frame rate of 30fps. Its semantic comprehension and visual rendering abilities have reached the state-of-the-art (SOTA) level among open-source models. When it comes to image-to-video generation, the model meticulously preserves the characteristics of the reference image, ensuring that the dynamic processes depicted adhere to the laws of physics. Additionally, the video extension feature can consistently generate extended videos up to 5 minutes in length, guaranteeing temporal coherence across frames and the rationality of physical motions. The model has been made available as open-source on platforms like GitHub and Hugging Face.