LongCat Avatar in ComfyUI does something a lot of you ask us about: take a photo, add a voice track and you get a video where the person talks, with the lips in sync. Meituan's version 1.6 came out on November 13 under the MIT license, and on November 18 Comfy-Org published the files ready for ComfyUI. There's no official node yet, but a community pack runs it right away.
What do we think? Until now talking avatars were mostly a job for paid online services like HeyGen, and online has real advantages: you upload, wait, download. But for a video tutorial, a presentation or an animated character who has to talk for two minutes, paying by the second gets expensive fast. We almost always work locally on an RTX 3060 with 12 GB, and we haven't tried LongCat 1.6 on that card yet. Here we lay out what's confirmed, which file to pick based on your video memory, and where the limits are.
| What it is | A model that generates videos of people talking from audio, a photo and/or text |
| Who made it | Meituan's LongCat team (China) |
| Video memory required | not stated; the model file ranges from 8.8 GB (W4A8) to 29.5 GB (bf16) |
| Does it run on an RTX 3060 12 GB? | not stated; on paper only the W4A8 version is the right size, but nobody has tested it yet |
| License | MIT (weights); the community node pack is GPL-3.0 |
| Commercial use | yes, the MIT license has no geographic or commercial limits |
| Where to download it | Hugging Face: Meituan's official repository and the Comfy-Org repack |
| Free or paid | free, it runs on your PC |
What is LongCat Video Avatar 1.6 and what's new?
LongCat Video Avatar 1.6 is an "audio-driven" video model: the voice decides how the mouth, face and body move. You can start from a photo plus audio, or just from text plus audio, and the video can keep going in chunks past 10 minutes. Compared with 1.5, the official card on Hugging Face lists three improvements:
- more natural lips, with less exaggerated mouth movement and less jitter from one frame to the next;
- more stable long videos, because errors don't pile up with every segment;
- faster generation: 8 steps in total, the first 4 with a "DMD" LoRA and the other 4 with an "OPSD" LoRA (two small files that speed up the model, which is why these are called distilled versions).
Meituan also published a preference test with 33,051 votes: LongCat 1.6 came second out of six, behind OmniHuman-2.0 and ahead of HeyGen, Kling Avatar 2.0 and its own version 1.5. The test was run by the people who made the model, so take it as a hint, not a verdict. The examples are on the project page.
How do you use LongCat Avatar in ComfyUI?
For now you use it with the community pack LongCat Avatar Native, which you install from the ComfyUI Manager. Official support is on its way: on November 18 Kijai opened a change to ComfyUI with a LongCatAvatarToVideo node that handles one to three speakers, but as of November 20 it hasn't been approved yet.
The pack adds four nodes: one loads the model, one turns the voice into data for the model (it uses Whisper, OpenAI's speech recognizer), and two generate the video. The first is the distilled 8-step version, the one you'll use almost always; the second is the 50-step base version, much slower, which also follows the text prompt. The video is built in blocks of 93 frames, and each block picks up from the last 13 of the previous one: that way the length follows the length of the voice.
If you've never installed a node pack, our beginner's guide to ComfyUI walks you through the Manager, the model folders and your first workflows step by step.
How much VRAM does LongCat Avatar 1.6 need?
Nobody has said, so you have to reason from the file sizes. The main model comes in four versions, which differ by quantization (the precision used to store the model's numbers: less precision, smaller file, slightly lower quality):
- bf16, the full version: 29.5 GB;
- int8: 14.9 GB, the one in Comfy-Org's official repack;
- W6A8: 12.3 GB;
- W4A8: 8.8 GB.
The last two are in a community mirror, not in the official repository. The pack's author measured the error against the full version: 1.4% for int8, 2.1% for W6A8 and 6.4% for W4A8. On a 12 GB graphics card like our RTX 3060, W4A8 is the only one that leaves some room; with 16 GB you can go for int8. You also need the umT5 text encoder (5.3 GB in int8), Whisper (about 1.2 GB) and the Wan 2.1 VAE, which ComfyUI can move to your PC's RAM when it isn't using them.
On speed there's only one figure, and it's on an RTX 4090 with 24 GB: at 768x512 about 5.1 seconds per step, and a 10.1-second video (253 frames) in about 162 seconds. On a 3060 it will take quite a bit longer. The pack uses 6.5–8 GB of system RAM, so with 32 GB you have plenty of room.
Can you use LongCat Avatar for work?
Yes: the weights are released under the MIT license, one of the most open there is, with no exclusions for Europe or any other region. It doesn't give you rights to Meituan's trademarks or patents, but there are no restrictions on the videos you generate. The node pack's code is GPL-3.0, which matters to people who modify and redistribute it, not to people who use it to make videos.
The real limit is something else: the photo and the voice have to be yours or belong to someone who gave you permission. Making a real person talk without consent is a legal problem, whatever the model's license says.
Try it now
- Update ComfyUI, search for "LongCat Avatar Native" in the Manager, then restart.
- Download the model in the version that fits your card (W4A8 for 12 GB, int8 for 16 GB or more) into models/diffusion_models, the two LoRAs into models/loras, umT5 into models/text_encoders, Whisper into models/audio_encoders and the Wan 2.1 VAE into models/vae.
- Prepare a clean voice track: the pack doesn't separate the voice from music, so no backing track underneath. If you only have a mixed file, separate the voice first with a dedicated program.
- Use a front-facing, well-lit photo that's already in the final format (landscape or portrait): a wrong crop is the most common cause of out-of-sync lips.
- Do your first test at 768x512 with 10 seconds of audio, using the distilled 8-step node. If it works, make the audio longer.
Our take: if you need a narrator or a talking character in a video, LongCat 1.6 is the first thing we'd try locally, mainly because of the MIT license. For scenes with two people, though, we'd wait for the official node.
If you notice small jumps in lighting between blocks, the node we covered in our article on how to remove flicker from AI videos with Deflicker can help. To compare LongCat with the other video models, including those that also generate audio, see our guide to AI video models in ComfyUI.
Other news
- ComfyUI 0.46.0 came out on November 18: it brings PAN 2, the Video Deflicker node and the Kling 4.0 Image nodes to the stable version. Heads up if you use an AMD card on Windows: DirectML support has been removed (release notes).
- PAN 2, MBZUAI's 29-billion-parameter "world model", now loads with ComfyUI's standard nodes. The int8 file weighs 26.7 GB, so it's for high-end cards (files on Hugging Face).
- MOSS-Image in 4 steps: on November 19 a community LoRA came out that takes MOSS-Image from 50 steps down to 4, under the Apache-2.0 license (model page).
Frequently asked questions
Does LongCat Avatar work with voices in other languages?
There's no official list of languages. The model reads the audio with Whisper large-v3, which recognizes many languages, and moves the lips based on the sounds, not on written words. In theory the language matters little, but nobody has published tests in languages like Italian yet: it's worth doing a test of a few seconds.
Can two people talk in the same scene?
The model can: Meituan provides scripts for multiple people, and the upcoming official node handles one to three speakers. The community pack, on the other hand, works with a single voice for now, because the two-speaker part hasn't been ported to ComfyUI yet. For two-person scenes it's better to wait for the official node.
How is it different from HeyGen?
HeyGen is an online service: you upload a photo and text, pay and get the video, no graphics card needed. LongCat runs on your PC, it's free and has no length limits, but you have to install everything and have enough VRAM. In Meituan's test LongCat 1.6 is preferred over HeyGen, but that comparison was done by Meituan.
What does W4A8 mean?
It's a type of quantization: the model's weights are stored in 4 bits and the intermediate calculations (the activations) in 8 bits. The file gets much smaller and fits on cards with less memory, at the cost of some precision. For LongCat the measured error is 6.4% compared with the full version.
If you want to know when local makes sense and when online is better, we cover it in our comparison AI locally or online. And on our YouTube channel @tutorial3d you'll find videos with free workflows to download.










