ComfyUI-MiniMaxH3-Prompt-Writer

ComfyUI MiniMax H3 Prompt Writer

<p align="center"> <img src="web/assets/h3-prompt-writer-launcher.svg" width="96" alt="H3 Prompt Writer"> </p>

Current release: 0.4.1 | MIT License

H3 Prompt Writer is a prompt-writing workspace for MiniMax H3 inside ComfyUI. Start with a plain-language Creative Brief, add optional image, video, or audio references, and generate an editable prompt in the format expected by H3.

It is a ComfyUI UI extension, not a workflow node. It writes prompt text for your existing H3 workflow. It does not run MiniMax H3, change the graph, or queue a video.

What's new in v0.4.0

• Added Qwen 3.8 support for Direct GGUF, including vision references and optional Thinking.
• Added support for compatible Qwen 3.8 fine-tunes and Qwen3-VL models.
• Reference assets can now be replaced without changing their position in the list.
• Improved automatic context selection for large prompts and prompts with many references.
• Improved Direct GGUF stability, memory handling, and tokenizer reliability.

What's new in v0.3

• Redesigned Writer and Settings interface.
• Ollama as a simpler local setup.
• Optional API providers.
• External llama.cpp now has its own dedicated provider setup.
• Saved drafts for every mode.
• Better automatic model and context handling.
• More reliable Reference prompts.

What it does

You do not need to write MiniMax section headings, timestamps, or reference syntax by hand. Describe the video and tell Writer what each reference should contribute:

Use <Picture 1> for character appearance, <Picture 2> for clothes, and only the movement from <Video 1>. The character walks through a rainy Tokyo street at night.

Writer sends your brief, selected mode, prepared references, and the official MiniMax prompt-writing guide to the chosen prompt model. The result is an editable H3 prompt. You can change it directly, use Refine for a revision, or select Copy prompt and paste it into your H3 workflow.

Key features

• T2VA, I2VA, FL2VA, L2VA, and Reference modes.
• Up to 9 images, 3 videos, and 3 audio references in Reference mode.
• Clear <Picture N>, <Video N>, and <Audio N> labels for assigning identity, wardrobe, setting, motion, camera, sound, or other roles.
• In-place Reference media replacement from the asset menu or by dropping one file directly on a card, without rebuilding the surrounding asset order.
• Ordered video contact sheets with visible frame-sampling controls, so you can inspect what the prompt model sees.
• Official MiniMax base and Reference guides included for all five modes.
• Editable prompts, Refine, Copy prompt, and a separate saved draft for every mode.
• Automatic context planning and clear controls for releasing local prompt models and ComfyUI VRAM.

See Writing a useful Creative Brief for practical examples.

Choose a provider

| Provider | Choose it when | Setup | | --- | --- | --- | | Ollama | You want the simplest local setup | Install Ollama and pull a vision model | | Direct GGUF | You want Writer to load a supported GGUF inside ComfyUI | Install the optional native runtime and add a matching GGUF + mmproj pair | | External llama.cpp | You already run llama.cpp or want full control over its runtime | Start llama-server; add a matching mmproj for images and video | | API providers | You want Gemini, OpenAI, OpenRouter, or a Custom OpenAI-compatible endpoint | Connect a key or an existing endpoint such as LM Studio |

Not sure? Start with Ollama. The provider guide explains the differences. The Ollama and Direct GGUF guides contain the tested local model choices.

Quick start

1. Install MiniMax H3 Prompt Writer from ComfyUI Manager and restart ComfyUI.

2. Open the floating H3 Prompt Writer button or use Extensions > H3 Prompt Writer. No graph node will appear.

3. Open Settings and choose a provider. For the recommended local setup and an 8 GB starting tier, install Ollama, open the app, and run:

   ollama pull gemma4:e4b
   

4. Choose a mode, add its media, and write a Creative Brief.

5. Select Generate prompt, review the editable result, then copy it into your H3 workflow.

For Git, ZIP, Windows Portable, update, and provider-specific steps, see Installation.

Privacy and limitations

• With Ollama, Direct GGUF, External llama.cpp, or a local Custom endpoint, the prompt request and prepared media stay on the local machine.
• With a remote API provider, the required brief, instructions, prepared images, and video contact sheets are sent to the selected provider. Original video and audio bytes are not uploaded by Writer. Read What leaves this computer before using private media.
• Video understanding uses the ordered contact sheet shown in the preview, not every frame of the encoded video.
• Prompt models do not listen to uploaded audio. Describe the soundtrack, voice, rhythm, or other audio role in the Creative Brief.
• The interface and documentation are in English. Briefs can use other languages, and Writer preserves supplied dialogue and visible text.
• Gemma 4 remains the simplest tested local choice. Direct GGUF also supports Qwen 3.8, compatible Qwen 3.8 fine-tunes, and Qwen3-VL. Untested compatible models may behave differently from the verified pairs.
• Ollama, External llama.cpp, and compatible API endpoints let you try other multimodal models that accept images. Compatibility does not guarantee a good H3 prompt.
• External llama.cpp also accepts text-only models for Music 3, T2VA, and Refine. Image and video requests still need a vision model.
• Direct GGUF supports Gemma 4, Qwen 3.8, compatible Qwen 3.8 fine-tunes, and Qwen3-VL. The tested Qwen 3.8 and Qwen3-VL model and projector pairs are marked as verified. Other compatible combinations are marked as unverified. A missing projector leaves text-only T2VA available.
• Gemini and a Custom OpenAI-compatible endpoint were tested live. OpenAI and OpenRouter have automated contract coverage but were not tested live with commercial credentials. Comfy Cloud has not been validated for v0.3.

Documentation

• Installation
• Using Prompt Writer
• Choose a provider
• Ollama
• Direct GGUF
• External llama.cpp
• API providers
• Troubleshooting
• Changelog

The project is released under the MIT License. MiniMax H3 guides and model files keep their upstream terms. Model weights are not bundled with this extension.