MiniMax References Manager

MiniMax H3 is the 33B omni-modal video model from the Hailuo folks - text, image, video and audio in one context, with native stereo sound - and it got day-zero ComfyUI support. The catch: the core MiniMax H3 Reference to Video node expects up to nine images, three videos, three soundtracks and three audio clips, each on its own socket, plus a prompt in a six-section format. Wire that by hand and you've built a plate of spaghetti you re-cook every time you swap a reference. This node removes that pain.

What it actually is

MiniMaxH3ReferencePack (display name "MiniMax References Manager") is a companion node, not a generator. It owns your whole reference set - you drop images, videos and audio into its UI instead of wiring loaders - and fans them out to the core node's 18 reference sockets. Wire it once and changing a reference never touches the graph. Empty slots emit None, which the core node skips, so the unused sockets just dangle.

It does two jobs beyond the plumbing. First, it labels every asset the way MiniMax will actually see it - <Picture 2>, <Video 1>, <Audio 1> - right on the tile. Second, and this is the party trick, it writes the prompt for you: a vision-language model reads your direction, looks at the references, and produces the six-section MiniMax prompt H3 expects. Two registers - standard writes a scene, replacement swaps something in a reference video for something in a reference image - and auto lets a cheap classifier pick.

It's emphatically not a wrapper - the source is explicit that it never touches H3's VAE or tokenizer. Generation still happens in MiniMax H3 Reference to Video - which is also why it downloads no model files; the whole dependency is requests.

The inputs that matter

The only required input is direction - your steer on subject, mood and action. Everything else has sane defaults. Three matter:

• prompt_provider - openrouter, local, or none; the whole architecture in one dropdown. openrouter calls a hosted multimodal model (needs a key). local points the writer at any OpenAI-compatible server on your machine - Ollama, LM Studio, llama.cpp, vLLM - no account, nothing leaves your box. none skips the API and passes your direction text straight through to prompt.
• reasoning_effort - none/low/medium/high, default medium. Only sent to OpenRouter, dropped for models that don't reason.
• max_reference_edge - downscales a reference image whose long edge exceeds it (default 2048, 0 off). Not cosmetic: MiniMax sizes references off their short edge, so a huge reference silently makes every sampling step pay for it.

There's a nice touch in local mode: the Local LLM button sweeps the usual local ports (LM Studio 1234, Ollama 11434, llama.cpp 8080, vLLM 8000), lists every server that answered, and fills in api_base and local_model_slug in one click. The scan is loopback-only.

The outputs

You get 20 sockets: image_1–image_9, video_1–video_3, video_audio_1–video_audio_3 (each video's soundtrack, extracted unless you toggle it off), audio_1–audio_3, plus prompt and debug. Wire the 18 reference sockets and prompt into MiniMax H3 Reference to Video and you're done. The debug socket dumps the whole request - where it posted, the model, every setting, each content part - worth a preview node.

Installing it

ComfyUI Manager, search "ComfyUI-MiniMaxRefPack". Or the manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/Hearmeman24/ComfyUI-MiniMaxRefPack
pip install -r ComfyUI-MiniMaxRefPack/requirements.txt

Then restart ComfyUI. A complete Reference-to-Video workflow ships with the pack (Workflow → Browse Templates → ComfyUI-MiniMaxRefPack), so you can see it all wired up first.

Where people get burned

• Local scan finds nothing. Start your server first (ollama serve, or LM Studio's Developer tab), then Rescan. And localhost means the machine ComfyUI runs on - if ComfyUI is in Docker or on a pod, its localhost isn't your laptop.
• Local prompts come out weaker. A local server takes text and images, not video or audio, so a clip goes as six still frames with no sound. The node says so on the canvas and in debug - honest, not broken.
• A key typed into the node lives inside the workflow JSON. Sharing that workflow ships your key - use OPENROUTER_API_KEY or LLM_KEY instead, which take precedence and stay out of the file. On local mode the environment is never read, so a stray key can't follow a pasted URL to someone else's server.
• Odd retention_analysis lines in the prompt. Known quirk: the packaged system prompt asks for one line per label, while MiniMax's guide says invented content gets none - the model sometimes resolves that by dropping or inventing a line. The debug socket shows exactly what it was told.

One thing to keep separate: this node manages references and writes prompts - running H3 still needs the core node and weights, which the MiniMax H3 Community License currently bars in the US, EU, UK and Korea. The node is MIT, needs no account, and only calls the provider you choose.