LongCat Avatar 1.6 in ComfyUI: Free Talking-Head Video From One Photo and One Voice Line
An open, MIT-licensed model now turns a portrait and a recorded line into a presenter who says it. Here is what it takes to run, what it costs in time, and where it falls short.
In the space of two days this week, a talking head became something you can get for nothing three different ways.
On Tuesday, Google made personal avatars in Vids free for anyone with a personal Google account. The same day, ComfyUI added paid nodes for HeyGen's Avatar 6. On Wednesday, a small outfit called PixelAI Labs published a free ComfyUI workflow for LongCat-Video-Avatar 1.6, an open model from the Chinese delivery company Meituan, and within hours ComfyUI's own team posted repackaged model files and began work on built-in support.
When a capability shows up free at the bottom of the market and as a paid service at the top in the same week, it has stopped being special. That is good news if you make explainers, course videos, character pieces or music promos. It also changes where your effort should go, and that is the real story here.
Of the three, LongCat is the one worth understanding properly, because it is the only one you can own.
What LongCat Avatar actually does
You give it three things: a photo of a person, an audio file of speech, and a sentence or two describing the scene. It gives you back a video of that person saying those words, with the lips, face and upper body moving in time with the voice.
The video is exactly as long as the audio. There is no clip limit to plan around. Meituan's model card says the model holds up in "videos exceeding 10 minutes," and version 1.6 claims better lip sync, more natural expression and better coordination between face and body than 1.5.
Three details matter more than the headline.
First, the license. LongCat is released under MIT, which is about as permissive as software licenses get. You can use it commercially, change it and build it into client work. There are no territory exclusions and no revenue thresholds, though the card notes it grants no rights to Meituan's trademarks or patents and expects you to follow privacy and content-safety law. That sets it apart from several other open video models of the past year, which shipped with licenses that excluded whole countries.
Second, it runs offline. Once the files are on your machine, nothing goes to a server. For anyone working with unreleased material or a client's face, that matters.
Third, it builds the video in short pieces. The PixelAI Labs workflow generates the first piece of about 3.7 seconds, then continues in pieces of about 3.2 seconds, each picking up where the last one ended. That is why long videos are possible, and it is also why time per second of output is the number you need to plan around.
Three ways in, all from the same week
Here is how the pieces fit together.
The PixelAI Labs nodes, released November 18, are the way in today. You install them through ComfyUI Manager, drag in a ready workflow file, and the loaders download what they need the first time you run it. The README is unusually clear for a first release: what to install, what each node does, which settings to leave alone, and real timings.
ComfyUI's own support is on the way. Comfy-Org posted repackaged files on Hugging Face on the evening of November 18, and the developer known as kijai opened a pull request that adds a built-in LongCat node with room for one to three speakers in the same shot. As of this morning that pull request has not been merged, so treat it as coming soon, not here.
A speed launcher called longcat-avatar-fast takes a different route. It runs Meituan's own scripts outside ComfyUI with a stack of speed tricks, and its author reports about three to four times faster rendering on the same card. It is for people comfortable with a command line on Windows, and its figures are the author's own.
If you live in ComfyUI and want to try this today, start with PixelAI Labs.
What it costs: time and hardware
There is no subscription and no per-clip fee. The cost is your graphics card and your patience.
Every published timing for LongCat comes from an RTX 5090, a top-end consumer card with 32 GB of memory, paired with 64 GB of system memory. On that machine, PixelAI Labs reports that a 3.8-second clip took between 5.6 and 7.7 minutes at 480p, depending on speed settings. Each additional 3.2 seconds or so of audio adds another 1.5 to 2.5 minutes.
Do the math for a real piece. A one-minute explainer is roughly 18 more pieces after the first, so expect somewhere between half an hour and nearly an hour on a 5090, before any retakes. The speed launcher's author reports a 3.2-second piece in about 30 seconds on the same card at its faster default setting (about 39 seconds at the same quality setting PixelAI Labs uses), which would bring a minute of video to roughly ten or eleven minutes including loading. That is a big difference, but it means leaving ComfyUI.
The part the release posts gloss over is memory. PixelAI Labs says plainly that it has not tested cards with less than 32 GB. Its notes mention that, with its default streaming setting, the model itself used about 13 GB on their card, and they suggest specific settings for lower memory. That is encouraging, not proof. If you have a 12 or 16 GB card, you are the tester.
The first download is also large: about 25 GB of model files that the workflow fetches automatically, plus a 6.7 GB text file you place yourself. Budget the disk space before you start.
If you do not own a card like this, rent one by the hour from a cloud GPU service. Check current hourly prices before you start, set up once, and batch your presenter clips into one session so you are not paying for setup time twice. The files you make are yours.
Put this into practice: your first session
Plan for one evening. Here is the order that wastes the least time.
- Install the nodes. In ComfyUI Manager, choose to install via Git URL and paste the PixelAI Labs repository address. Restart ComfyUI.
- Add the text file. Download the
umt5_xxl_fp8_e4m3fn_scaledfile linked in the README and put it in yourmodels/text_encodersfolder. The workflow fetches everything else on the first run. - Load the workflow. Drag
LongCat_Avatar_1.6_PixelAILabs.jsonfrom the repository's example folder into ComfyUI. - Pick your photo carefully. Chest up, framed exactly the way you want the final shot. The model animates what it is given; it does not reframe.
- Record or generate clean speech. No music, no room echo. The README's examples use voice lines from ElevenLabs, but your own recorded voice works the same way. If your only copy has music under it, the workflow includes an optional node that pulls out the voice.
- Write a short scene description. Who the person is, where they are, and the phrase "static camera on a tripod." If you want a calm presenter, add "hands relaxed" or "hands out of frame."
- Test with four seconds of audio first. That is one piece, a few minutes of waiting, and it tells you whether your photo and voice work before you commit to a long render.
- When a take looks wrong, change the seed, not the prompt. The seed is the number that controls the random variation in each run. The README's tips say exactly this, and it saves a lot of rewording.
Leave the resolution at 480p while you learn. The README says 720p is about three times slower per step.
Where it falls short
Here is the honest list.
It is new in ComfyUI. The PixelAI Labs nodes are a day old, with a handful of stars on GitHub. They are built on an earlier node pack for version 1.5, which helps, but expect rough edges and updates.
Smaller cards are unproven. Everything you read about speed comes from one very expensive card. Until someone publishes a result on 16 or 24 GB, assume you are experimenting.
480p is the practical ceiling for now. That is fine for a phone-screen explainer or a social clip. It is not a hero shot. You may want to run the result through an upscaler, which adds another step and another chance for faces to drift.
It animates; it does not direct. You get a person talking to camera. You do not get blocking, cutaways or camera moves. A good edit still needs B-roll, and the presenter only holds attention as long as the words do.
Languages are not well documented. The model card tags English and Chinese. If you need other languages, test with a short clip before you plan a project around it.
There is no consent check. This one matters most. Google limits Vids avatars to the account holder's own face. LongCat has no such limit, because MIT licenses do not police use. You can make anyone appear to say anything from one photograph. If you are making avatars of clients, collaborators or anyone who is not you, get written permission for both the face and the voice before you render a frame, and keep it on file.
Which one should you actually use?
If you need one talking explainer of yourself this week and own no graphics card, use Google Vids. It is free, it is fast, and the consent rules are built in.
If you want HeyGen's look inside a ComfyUI graph that also handles your backgrounds and color, the new HeyGen nodes are the paid path.
If you want something you own, can run without an internet connection, can use for client work without a revenue clause, and can extend to long videos, LongCat is the one. The price is hardware and an evening of setup.
The scarce part moved
For years the hard part of an AI presenter was getting the mouth to move convincingly. That problem is now solved three ways, at three prices, in one week. A presenter who says something dull in perfect sync is still dull.
So spend your effort where it now counts. Write the script as if a camera crew were costing you money. Record the voice take until it sounds like a person who means it. Pick the photo the way a director picks a lead. LongCat will handle the lips.
Try one four-second test tonight. If your photo and your voice survive it, you have a presenter you own outright, and the only thing left to get right is what they say.