Chat with a character in ComfyUI; each reply is a directed MiniMax H3 video in that character's look and voice, rendered on your own GPU.
How Chat2Me Works: From First Message to Film. (Click to watch on YouTube)
This repo is the front door: this README and the issue tracker. There is no code here. Chat2Me's code is in a private repo that sponsors get access to.
- Make a character from a RefCap or from one photo of a face, and give them a voice.
- Type to them. A director LLM (Ollama, or the claude CLI) turns your message into up to three shots, and ComfyUI renders them as MiniMax H3 clips with speech.
- Keep a story: worlds with places, a cast, an editable story ledger, continuity between clips, branches and retakes.
- Publish a session as one film, with title cards, fades, credits and subtitles.
One session, in the order it was made.
 Elowen and Brokk in one reply, each in their own voice. |
 Vesper, made from one photo of her face, in her tower study. |
A character is a RefCap's face views, or one photo of a face, plus a body RefCap that carries the outfit and a voice reference. Vesper started from her photo; the director wrote her speech style.
Hollowsong has two places with sets of their own, a crystal cavern and a tower study, and a start: the place, the hour, a prop, an open thread and the first message.
The director plans the shots, ComfyUI renders them, and the card shows how far it has got. Clips play back to back; Show words lists what is said and done.
Take a picture of a frame, crop the part you mean, name it, and put it on your next message. The reply holds it.
All of it comes from one session at Medium quality (960x544) on an RTX 4070 12 GB, rendered with a community MiniMax H3 turbo fine-tune and directed by a Gemma 4 12B variant in Ollama. The characters and places are made up: faces from a community Krea 2 fine-tune, bodies and sets rendered with H3, voices designed with VoxCPM2; no one's face or voice was used. They are free to download as RefCaps on Hugging Face.
Sponsor on GitHub for US$5 a month. GitHub then invites you to the
private repo aemotion-studio/ComfyUI-Chat2Me; clone it into ComfyUI/custom_nodes.
Chat2Me is Sold Open v0.2: you pay for access, not for permission.
- What the price pays for: access to the private repo that holds Chat2Me, every version as it is released, and direct support. Cancel any time.
- Licence: every version is GPL-3.0-or-later.
- Source: the code is the download, in the sponsors' private repo. It is not public.
- Early access: none. Every sponsor gets every version the day it is released.
- Sharing: you may share what you get. Nothing checks whether you paid, and when a sponsorship ends, what you downloaded stays yours.
- Other terms: the models below are not part of Chat2Me and carry their own licences.
- ComfyUI with its native MiniMax H3 nodes, and ComfyUI-RefCap (free).
- The H3 files from Comfy-Org/MiniMax-H3, which you download yourself under the MiniMax H3 Community License. As of October 2026 it does not cover the United States, the European Union, the United Kingdom or South Korea; read it before you download.
- A director: Ollama (
gemma4:12bby default), LM Studio or another OpenAI-compatible server (Unsloth Studio, llama.cpp, vLLM), or theclaudeCLI, which sends the story to Anthropic and bills your account. ffmpeg, for publishing a film.
Built and measured on one machine: an RTX 4070 12 GB with 31 GB of RAM, on Linux. There, on a community H3 turbo fine-tune, a 10 s clip at Low quality took about 86 to 160 s. One turn or build runs at a time, and there is no access control, so keep it on your own machine.
Open an issue here.