Meituan's LongCat team has upgraded its open-source digital human model, LongCat-Video-Avatar, to version 1.6. The weights are on Hugging Face under the MIT license, so commercial use is allowed too. This release adds no flashy features. It mainly fixes the two problems that annoy people most when making talking videos: the lips look fake, and the person starts to drift away from their look when the video gets long.
What it can do
Give it an audio clip of someone speaking, and it generates a video of a person whose lips move with the sound. It supports these uses:
| Use | What you give it | Good for |
|---|---|---|
| Audio + text description | A voice track, plus a sentence about what kind of person and what setting | No photo of a person, and you want a presenter generated from scratch |
| Audio + person image | A voice track and a portrait | A series of talking videos with the same look |
| Continuation | More footage generated after an existing clip | Videos several minutes or longer |
| Multiple people | A separate audio track for each person | Dialogue and interview videos |
What changed in 1.6
According to the official model page, there are three changes from 1.5:
- More natural lips: Before, the mouth opened too wide, the lips looked unnaturally red, and the mouth area flickered. This version reduces all three. The movements of the face, head, and body also fit together better.
- Long videos hold together: Long videos are generated one segment at a time, each one building on the last. Before, the error grew a little with every segment, so the person looked less like themselves toward the end. This version keeps that buildup under control. The official claim is that it can make 10 minutes or more, and the person still looks like the same person.
- Faster generation: The whole pipeline has been sped up. There is also an optional INT8 compressed version of the weights that takes up less space.
The official team also ran a round of human voting comparisons against commercial products, including OmniHuman-2.0, HeyGen, and Kling Avatar 2.0. 1.6 ranked second, behind only OmniHuman-2.0 and ahead of HeyGen and Kling. Meituan organized this test itself, so treat it as a reference only.
Can a regular computer run it
The official examples run on two graphics cards working together. The setup needs PyTorch, FlashAttention, and similar tools, which are hard on home graphics cards. A more realistic route on your own card is ComfyUI. The community plugin WanVideoWrapper already supports the previous version, 1.5. People have also shared compressed versions such as fp8 and GGUF. When video memory is short, you can use block swap to move part of the weights into system memory. ComfyUI support for 1.6 is not yet verified, and it may have to wait for a plugin update.
How to get started
If you don't want to set up the environment yourself, give the prompt below to your AI agent. It can first check whether your computer can run the model and which route to take.
I want to try Meituan's open-source digital human model LongCat-Video-Avatar 1.6 (Hugging Face: meituan-longcat/LongCat-Video-Avatar-1.6), and generate a talking video from one audio clip and one portrait image. Please help me:
1. First check my computer's graphics card model and video memory, and tell me whether I can run it the official way or whether I have to use ComfyUI
2. Check the model page and the WanVideoWrapper repository to confirm whether ComfyUI supports 1.6 right now. If it doesn't, tell me to try 1.5 first
3. Based on the route you pick, list the files to download and roughly how much disk space they take. Wait for my confirmation before downloading
4. After installation, run one clip using an audio file of 10 seconds or less, and tell me how long it took and how much video memory it used
5. If something requires installing drivers, paying, or logging into an account, stop and ask me
Show me your assessment and plan first. Don't start installing yet.
Start with a short audio clip of 10 seconds or less. First runs of this kind of model most often get stuck on the environment or video memory. A short clip fails fast and is easier to debug. Once it runs, move on to longer videos.
Notes
- Get permission for portraits: Using someone else's photo or voice for a digital human may infringe their portrait and voice rights. Also label the video as AI-generated when you publish it, as each platform requires.
- Treat the results as official claims: The improvements to lip movement and long-video stability, and the rankings, all come from Meituan's own model page. Run a test yourself to see the real results.
- Large downloads and disk use: You need two sets of weights, the base model LongCat-Video and Avatar 1.6. Leave enough disk space first.
For the idea behind making a short digital human video from one sentence with a commercial model, read this experience.


Comments (0)
No comments yet. Start the conversation.