MiniMax H3 Studio 👻

The pack's H3 story is a pipeline: Optimizer writes the prompt, Unpacker splits the pipe, and ComfyUI's official MiniMax H3 nodes turn that into conditioning. MiniMax H3 Studio 👻 collapses all three into one node. You attach your materials, describe the video, hit run - and out come the positive conditioning and latent ready for an H3 sampler, plus the generated prompt so you can see what it actually did.

It's the node the README actually advertises ("Minimax H3 Studio 海螺H3工作台"), and it's the reason most people install this pack. One workflow covers text-to-video, image-to-video (with or without a last frame), and the "all-reference" mode with up to 9 images, 3 videos and 3 audio tracks.

The inputs, in order of importance

• clip and vae - the H3 text encoder and video VAE, from the official MiniMax H3 loaders.
• h3_materials - the same upload widget as the Optimizer: thumbnails, click-to-insert @图1 / @视频1 / @音频1 references. This is where your frames, reference videos and audio go.
• mode - T2VA / I2VA / FL2VA / L2VA / Ref2VA, the same five-way switch that decides how many images the widget accepts and how keyframes are routed.
• style_preset - 13 presets feeding the prompt-writing system prompt.
• duration - seconds (1–15), auto-converted to the frame count H3 wants.
• width / height - default 1344×768, the H3-friendly ballpark.
• audio_vae (optional) - required if Ref2VA mode has any audio material; the node refuses to run without it, which is the correct behavior.
• ref_image_size (optional) - only matters in Ref2VA: match fits references to the generation resolution, max keeps them at full fidelity (2048 short edge).
• llm_config (optional) - the LLM Config node; without it, prompts pass through unrewritten (still valid, just not LLM-enhanced).

Outputs: positive (CONDITIONING), latent (LATENT), prompt (STRING), prompt_zh (STRING). Wire the first two into a MiniMax H3 sampler node and you're generating.

How it works

Under the hood it's honest glue. It runs the same prompt pipeline as the Optimizer (LLM rewrite of your idea into H3-grammar, tag resolution, hallucination scrubbing, Chinese review copy), then converts your duration in seconds to frames on H3's odd "17k+5" frame grid (H3 wants clip lengths of 22, 39, 56, 73... frames - durations that land on n ≡ 5 mod 17 at 24fps), then calls ComfyUI's official MiniMaxH3ImageToVideo or MiniMaxH3ReferenceToVideo node depending on mode. Your positive conditioning and latent are literally the official node's output - this node is the front-end that feeds it.

The mode routing is where the value is. In I2VA your image becomes first_frame; in FL2VA the second image becomes last_frame; in Ref2VA the uploaded materials are split into the right ref_image / ref_video / ref_audio slots automatically. No wiring mistakes, no wrong socket, no "wait, which port does the reference video go into."

Install & deps

# ComfyUI Manager: search "ComfyUI-Minitools"
# or manually:
cd ComfyUI/custom_nodes
git clone https://github.com/lepiai/ComfyUI-Minitools
# restart ComfyUI
pip install openai

The big requirement is a ComfyUI version that ships the official MiniMax H3 nodes - this node imports MiniMaxH3ImageToVideo and MiniMaxH3ReferenceToVideo from core and hard-fails with a clear message if they're missing. ffmpeg (with image2) for video references, PyAV/torchaudio for audio.

The things to know before you commit

• The H3 model licence is geographic. MiniMax opened H3's weights in August 2026, but the community licence excludes the US, EU, UK and Korea - users there aren't licensed to run the local weights at all (the hosted Hailuo API is a separate, globally available thing). This node generates conditioning for H3, so if you can't legally run H3, the node's usefulness starts and ends at the prompt output. Check before you build a workflow around it.
• It's a cloud-LLM node for the prompt half. Your description and materials hit the configured LLM API. No key connected → offline passthrough mode, still generates (with whatever prompt you typed).
• Ref2VA + audio is a hard requirement, on purpose. Forget the audio_vae and it refuses to run rather than silently producing garbage. That's a feature.
• Default model strings are dated. The default LLM config points at a specific Qwen snapshot on Aliyun's Bailian; if the endpoint 404s later, update the model name.

If you're in the MiniMax H3 world and legally able to run it, this is the single node that turns a stack of official ComfyUI parts into a clean, one-node video studio. If you're not, it's still a surprisingly complete H3 prompt-and-package tool - you just won't use the positive and latent outputs.