MiniMax H3 Companion — Interactive Talking Avatar (local LLM, no API)
Alternatives:
MiniMax H3 Companion — ask a question, get a video answer
Talking Avatar Companion (MiniMax H3 + Ollama)
Tags
workflow comfyui minimax minimax-h3 video talking-head avatar ollama llm text-to-video voice-clone subgraph
Description
Ask a question, get a video of your avatar answering it. Out loud, in its own voice, with its own face.
A local LLM writes the answer and stages it as a shot; MiniMax H3 renders the shot with native synced audio. Nothing leaves your machine — no API keys, no cloud calls.
How it works
You type a question.
Ollama sends it to a local vision model along with the avatar image, so the avatar can see what itself looks like.
The LLM replies in JSON: a scene description of itself speaking, plus how many seconds it needs.
The scene description becomes the MiniMax H3 prompt; the duration is converted to a valid frame count.
MiniMax H3 renders the avatar speaking, with the voice cloned from your audio sample.
The conversation is remembered between runs, so you can keep talking to it.
What you get
Everything lives inside one Companion Engine subgraph. On the canvas you only touch:
AvatarThe face. Doubles as <Picture 1> for MiniMax and as the image the LLM sees.Voice sampleThe voice to clone (<Audio 1>). A few seconds of clean speech is enough.Ollama serverURL and model. Defaults to a local Ollama at 127.0.0.1:11434.Your questionThe only field you change between runs.ResolutionOutput size, with a size reference note on the canvas.Four loadersDiffusion model, text encoder, video VAE, audio VAE.
Three outputs: the video, the LLM's raw JSON, and the exact prompt MiniMax received — the last two make debugging a bad take trivial.
Be aware that the videos take time to generate, my companion is in the outer rim of the solar system and its answers take some time to reach.
Requirements
ComfyUI, recent enough for
MiniMaxH3ReferenceToVideoand subgraph support.comfyui-ollama — the only custom node pack needed. Available in the ComfyUI Manager.
Ollama running with a vision-capable model pulled. The avatar has to be able to see its own picture, so a text-only model will hallucinate its appearance.
Every other node is ComfyUI core. No KJNodes, no Easy-Use, no attention patches, no cache nodes, no upscalers. That is deliberate: it should load for everyone and be easy to extend with whatever speed tricks your hardware likes. I'd suggest you apply as many optimisations as you can, but as these change rapidly I decided not to include them.
You need to be able to run Ollama and Minimax H3, so the VRAM+RAM requirements might be steep. I am running with 16GB VRAM + 64 RAM and each answer takes ten minutes or so to generate for me.
Models
From Comfy-Org/MiniMax-H3:
ComfyUI/models/
├── diffusion_models/
│ └── minimax_h3_ref2va_pruned_int8_convrot.safetensors
├── text_encoders/
│ └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
└── vae/
├── minimax_h3_video_vae_fp16.safetensors
└── minimax_h3_audio_vae_fp32.safetensorsNote this is the ref2va checkpoint, not the fl2va one used by the t2v/i2v templates.
Setup
Install comfyui-ollama and download the four model files above.
Start Ollama and pull a vision model.
Load the workflow. In the Ollama server node, hit refresh and pick your model — the field ships empty on purpose.
Load your avatar image and a voice sample.
Type a question and queue it.
Notes and tips
Memory:
keep_contextis on inside the subgraph. The avatar remembers previous turns for as long as ComfyUI stays open. The context window is not very big.The LLM decides the length. It reports how many seconds it needs and a Math Expression converts that to frames. MiniMax H3 only accepts lengths where
length % 17 == 5— 5 s is 124 frames, 10 s is 243, 15 s is 362. The system prompt caps it at 10 s; raise it if you want longer takes.Structured output, not regex. Ollama runs in JSON mode and the keys are pulled with
Extract Text from JSON. Parsing free text here is a trap: the duration line ends up inside the video prompt and MiniMax tries to render it.Bad take? Check the Answer (raw JSON) output first. Most misfires are the LLM drifting from the format, and that is fixable by editing the system prompt inside the subgraph.
Prompt wording matters a lot with ref2va. Keep the
<Picture 1>and<Audio 1>tags and be explicit about which reference drives what.Upgrade path: if your comfyui-ollama is recent, the
Ollama Chatnode handles history with proper roles and has areset_sessiontoggle. Drop-in improvement overOllama Generatefor this use case.
Credits
Built on the official ComfyUI MiniMax H3 ref2va template. LLM nodes by stavsap.
Description
A more advanced interaction thanks to a new continuity mode: the last frame of the video is isolated and fed to the model.
This is not by any means a product but a concept-prototype.
It uses ollama, but if you don't like ollama you probably can figure how to change that yourself. My experience with LLMs is limited.
This is how you play the game:
First round. An image, a setting, a voice, and two prompts one for the LLM and one that only Minimax will read. There are two toggles Continuity and Reset Memory. Continuity to false and Reset Memory to True. Make sure that the Continuity nodes can read an image even if it will not be used.
Hit run and wait for the round to complete. In my 4060 Ti with the backup of a macbook M1 executing the LLM every round is about 4 minutes.
Second round and on: change both toggles (Continuity true, Reset Memory false). Edit the conversation prompt and hit run.
You can experiment with the prompt, I use the prompt for the LLM to tell the avatar who she is, her personality and goals, and my questions. The hidden prompt you can use to create things that the companion isn't aware of, unless it sees it in the last frame when the next round comes.