Introducing LongCat-Video: Meituan Releases Unified Video Generation Model, Challenging the Limits of Minute-Long Videos

Explore LongCat-Video, the latest AI video generation model released by Meituan. It is not only a unified framework that can handle various tasks such as text-to-video and image-to-video, but it also excels at generating high-quality videos up to several minutes long, taking an important step towards a “world model”.

The AI video generation track has been getting more and more lively recently. While we were still marveling at the amazing effects demonstrated by OpenAI Sora or Kuaishou Kling, another heavyweight player has joined the race with unique technology.

That is LongCat-Video, a unified basic video generation model released by the Meituan team.

You may be thinking, another AI video tool? What’s so special about it? To be honest, it does have a few very attractive highlights, especially in solving some of the core pain points of current AI video generation.

Not Just a Single Function, This is an “All-in-One” Unified Model

Many AI models focus on a single task, such as “text-to-video” or “image-to-video”. But LongCat-Video takes a more integrated approach. It adopts a unified architecture that integrates multiple mainstream video generation tasks into one model.

This means that whether you want to:

• Text-to-Video: Input a text description to generate a corresponding video.
• Image-to-Video: Given a static image, make it move.
• Video-Continuation: Continue an existing video to generate subsequent content.

LongCat-Video can handle them all with the same core model. This is like having a Swiss army knife for video creation, rather than a bunch of separate tools, which greatly simplifies the workflow.

The Real Highlight: Efficiently Generating “Minute-Long” Videos

This is probably the most exciting feature of LongCat-Video.

If you have played with other AI video tools, you may have found that it is easy to generate short clips of a few seconds, but it is a huge challenge to produce a video that is several minutes long, with coherent content and stable image quality. Many models, when the time is extended, will experience problems such as drastic changes in screen style, color drifting, or inconsistent characters, like a storyteller who forgets what the protagonist looks like halfway through the story.

LongCat-Video cleverly solves this problem. Its secret weapon is that the model focuses on the “Video-Continuation” task during the pre-training phase. In other words, it was trained from the beginning to be a master of “story relay”.

This native continuation ability allows it to better maintain content coherence and quality stability when generating long videos, avoiding problems such as screen collapse or style confusion. According to the official demonstration, it can produce videos up to several minutes long without significant quality degradation.

How Does It Do It? A Glimpse into the Technical Magic Behind It

It sounds amazing, right? The high efficiency and high quality of LongCat-Video are mainly due to the combination of several key technologies:

• Coarse-to-Fine Generation: This method is very intuitive, just like a painter painting, first making a draft, and then gradually improving the details. The model will first generate a low-resolution video prototype, and then gradually improve the resolution and details, and finally produce a 720p, 30fps high-quality video. This not only improves efficiency, but also ensures the final quality.

• Block Sparse Attention: This is a clever design to improve computing efficiency. The traditional attention mechanism allows the AI to process all the information on the screen at once, which is very resource-intensive. Block sparse attention allows the AI to “focus” on the most important parts of the screen and skip irrelevant areas, which is both smart and labor-saving, and greatly speeds up the generation speed.

• Multi-Reward RLHF: You may have heard of RLHF (Reinforcement Learning from Human Feedback), which is to let the model learn from human preferences. LongCat-Video goes a step further and adopts a “multi-reward” mechanism. This means that it not only learns “whether it looks alike”, but also judges the quality of the video from multiple dimensions, such as: screen aesthetics, action fluency, story logic, and fit with the text description. This makes the final video more in line with human aesthetics and expectations.

Not Just Generation, But Also “Interactive” Video Creation

LongCat-Video also demonstrated a very interesting feature: interactive video generation.

This means that users can intervene and give new instructions during the video generation process like a director. For example, you can first generate a scene of “a girl cutting bread in the kitchen”, and then when the video continues, input a new instruction “she pours a glass of milk”, and the model will seamlessly generate the next action.

This ability allows creators to no longer be passive recipients, but active participants who can guide the direction of the story, bringing unprecedented freedom and imagination to video creation.

Want to Try It Yourself or Learn More?

The Meituan team has been very generous in open-sourcing the resources related to LongCat-Video, so that everyone can access this technology.

• Official Website: https://meituan-longcat.github.io/LongCat-Video/
• GitHub Code: https://github.com/meituan-longcat/LongCat-Video
• Hugging Face Model: https://huggingface.co/meituan-longcat/LongCat-Video

Interested developers or creators may wish to go to the official page to see more amazing demonstration videos, or go directly to GitHub and Hugging Face to download the model and code to experience it for themselves.

A Small Step Towards a “World Model”

In summary, LongCat-Video is not only a powerful AI video generation tool, but it has also made important progress in the two key directions of “unified architecture” and “long video generation”.

The official positioning of it is “our first step towards a world model”. The so-called “world model” refers to an AI system that can understand and simulate the operating laws of the real world. And the ability to generate coherent, long-sequence videos is the basis for simulating the dynamic changes of the world. From this perspective, LongCat-Video does show great potential, and it also makes us full of more imagination for the future of AI.