MiniMax H3 Prompt Writer
Choose a MiniMax H3 mode, add your photo or other references, and describe the video you want in plain language. Writer uses your selected LLM to turn that brief into a prompt formatted for MiniMax H3. Review it, make any changes, and copy it into your video tool or workflow.
ComfyUI · v0.4.6
Runs inside ComfyUI. Includes workflow media integration and Auto VRAM.
Windows Standalone · v0.1.7
Runs without ComfyUI. Includes the same prompt writing, Sequence, Composer and Editor tools.
Select the version for your setup on this Civitai page and download its ZIP. The two editions use separate version numbers.
What's New
In both editions
Sequence: write prompts for longer videos by splitting one brief into timed chunks. Each chunk gets its own complete, editable prompt, with shared references and individual refinement. Copy the prompts into your preferred video tool or use them in a manual workflow. Choose Official for the H3 prompt structure or Compact for descriptive prompts that are easier to adapt to other tools.
Media Composer: combine pictures and video contact sheets into a reference collage.
Media Editor: crop pictures, trim and crop video, and extract frames.
Light theme and interface size: switch themes and make the UI larger.
ComfyUI only
Add to workflow: drag media from the floating panel onto the canvas to create a loader, or onto a compatible loader to replace its file. Supports native image, video and audio loaders, plus VHS Load Video (Upload). Drag again to transfer later edits.
Auto VRAM: coordinate memory between ComfyUI and supported local Writer models.
Core Features
Supports all current H3 input modes:
T2VA
I2VA
FL2VA
L2VA
Reference
Reference mode supports up to 9 images, 3 videos and 3 audio references, with 12 files total.
Tell Writer what each reference should provide:
Use Picture 1 for the character, Picture 2 only for the clothes, and only the movement from Video 1. Put the character on a rainy street at night.
Writer handles the H3-specific prompt structure and reference syntax around that.
Other features:
Editable generated prompts
Refine an existing prompt with a short instruction
Separate saved drafts for every H3 mode
Insert Picture / Video / Audio tags directly into the text
Video contact sheets with frame sampling controls
Automatic context handling
Local model unload and VRAM controls
Custom system prompts
Official MiniMax prompt-writing guides used during generation
MiniMax Music 3 workspace
Writer prepares prompts and reference media. It does not generate videos. Copy the finished prompt into your H3 tool or workflow.
Prompt Model Providers
The model writing the prompt is separate from the MiniMax H3 model used in your video workflow.
There are currently four ways to run it.
Ollama
The simplest local starting point for most users. See the provider guide for tested models and setup.
GGUF
ComfyUI: Direct GGUF loads the prompt model through llama-cpp-python. Supports Gemma 4, Qwen 3.8, compatible Qwen 3.8 fine-tunes and Qwen3-VL. Images and video require a matching vision projector.
Windows Standalone: Local GGUF uses your own llama-server executable and existing GGUF models. Select the runtime, model and matching vision projector in Settings.
External llama.cpp
Connect Writer to your own llama-server.
Useful if you already use llama.cpp or want full control over model loading, GPU placement, context, KV cache and the runtime itself.
API Providers
Supports:
Gemini
OpenAI
OpenRouter
Custom OpenAI-compatible endpoints
Custom endpoints can also be used with local servers such as LM Studio.
When the provider runs on your computer, the prompt request and prepared media stay local. Remote providers, including a remote Ollama host, receive the data needed for the request.
Provider guide:
https://github.com/duckyshell/ComfyUI-MiniMaxH3-Prompt-Writer/blob/main/docs/PROVIDERS.md
MiniMax Music 3
Prompt Writer also includes a separate workspace for MiniMax Music 3.
Describe the track you want in normal language and Writer builds the structured Music 3 caption.
Lyrics can be written manually, generated from the Music Brief, or rewritten using Refine.
Music 3 is separate from H3 video prompting, but uses the same provider system.
Basic Usage
Open H3 Prompt Writer
Go to Settings and choose your provider / model
Select the H3 mode
Add your references
Write the Creative Brief normally
Press Generate prompt
Edit it, use Refine, or copy it into your H3 workflow
You do not need to manually write H3 timestamps, sections or reference formatting.
Installation
Select ComfyUI v0.4.6 or Windows v0.1.7 on this Civitai page and download the matching ZIP.
ComfyUI
Extract the ComfyUI package into ComfyUI/custom_nodes/.
The final folder should be:
ComfyUI/custom_nodes/ComfyUI-MiniMaxH3-Prompt-Writer/Restart ComfyUI. Open the floating H3 Prompt Writer button or Extensions > H3 Prompt Writer.
This is a UI extension. No Writer node appears in node search.
Installation guide
If you can't find H3 Prompt Writer in ComfyUI, open it from the Extensions menu or use the H3 Writer button:
Windows Standalone
Extract the Standalone package to a writable folder and run start.bat.
Requires Windows 10/11 x64 and Python 3.10+ or uv. The first launch creates its own Python environment. Models and llama.cpp binaries are installed separately.
GGUF Setup
For Direct GGUF inside ComfyUI, follow the Direct GGUF guide.
For Local GGUF in Windows Standalone, follow the Standalone setup guide.
Notes
Video references are shown to the prompt model as ordered contact sheets instead of the full encoded video.
Prompt models do not listen to uploaded audio files. Describe what should be taken from the audio, such as voice, music, rhythm or soundtrack.
The base ComfyUI extension has no additional Python dependencies. Provider setup is separate. Standalone installs its own dependencies on first launch.
The project itself is released under the MIT License.
MiniMax-provided guides and model files retain their upstream terms.
Links
GitHub
https://github.com/duckyshell/ComfyUI-MiniMaxH3-Prompt-Writer
Usage / Creative Brief examples
https://github.com/duckyshell/ComfyUI-MiniMaxH3-Prompt-Writer/blob/main/docs/USAGE.md
Troubleshooting
https://github.com/duckyshell/ComfyUI-MiniMaxH3-Prompt-Writer/blob/main/docs/TROUBLESHOOTING.md
Description
Added Qwen 3.8 and Qwen3-VL support
FAQ
Comments (12)
Hi, thank you for creating and maintaining this great extension.
I’ve been enjoying using it with MiniMax H3.
Would you consider adding support for a remote Ollama endpoint?
My setup is:
- ComfyUI runs on a Windows PC
- Ollama runs on a separate Linux PC on the same LAN
- I’d like to use the Linux GPU for the prompt model so that the ComfyUI GPU can remain fully available for MiniMax H3
At the moment, the Ollama provider seems to detect only a local Ollama instance.
I also tried connecting to the remote Ollama server through the Custom OpenAI-compatible provider using:
The connection itself works and the model is detected, but prompt generation repeatedly ends with GENERATION_TRUNCATED. It seems this may be related to differences in reasoning/thinking handling through the Custom provider.
If possible, it would be very helpful if the Ollama provider could allow a custom host URL, for example:
This would make it possible to run the prompt model on a separate GPU machine while keeping the ComfyUI GPU free for H3 generation.
Of course, I understand this may not be a priority, but I thought this setup might also be useful for other users with multiple machines.
Thank you again for the excellent work!
Glad you’re enjoying it!
Remote Ollama support is now available. Just update from GitHub and use “Use Ollama on another computer” in Ollama mode.
For GENERATION_TRUNCATED, my guess is that the model is spending too much of its output budget on thinking. Native Ollama mode lets Prompt Writer control thinking directly. If it still overthinks, try setting the model’s reasoning effort to low on the Ollama side, if supported.
Thank you so much for the quick update!
I updated from GitHub and tested the new remote Ollama option. It works perfectly with Ollama running on my separate Linux machine.
I’m now using Gemma 4 12B remotely with a 16K context, and prompt generation completes successfully without the GENERATION_TRUNCATED issue.
This is exactly the setup I was hoping for, and it lets me keep the GPU on my ComfyUI machine completely free for MiniMax H3 generation.
I really appreciate you taking the time to add this feature so quickly.
Thank you again for the excellent work!
I've been using this a lot lately, and its really helped me create scenes closer to my initial vision, Thank you for this!
one item i would like to ask, is are you able to split a prompt.
i use this workflow MiniMax H3 Continuum, and it has the ability to [Chunk} a scene, or i guess extend it in a way where the details of the character/environment are retained but the story can go on longer.
if i use H3 prompt to write a 20s or 30s prompt, can i have it split that into 2 consistent and connecting scenes? will i have to create 2 prompts and hope they match?
sorry if im not explaining it well.
i want to be able to write a 30s prompt or even a 60s prompt then have the H3 prompt write allow me to split it into chunks that i can feed into the workflow, or any extended workflow for that matter.
in either case, great work here!
You can copy the prompt from the previous chunk into the Creative Brief, add instructions for how the next segment should continue, or use Refine to adapt it for the next chunk. It won't guarantee perfect continuity, though. There is no dedicated Continuum support in the Writer right now.
@useruser00003 Thank you for the reply, i placed a long scene in the writer and then used the output to placed 10s [chunk] markers throughout the script in my workflow, it comes out surprisingly well, later chunks do forget some things, like "hair is wet" or "with this in hand" etc.. so i just add those back in. the below is an example of how i got a consistent 30 seconds clip
hard to explain but below is an example outline
All these fields stay at the top of the prompt all from whatever H3 writer spits out
subject_definitions: whatever H3 writer spits out - <subject 1> is character, <subject 2> is grabby thing, <subject 3> is wet location
summary: whatever H3 writer spits out - character at wet place grabs the thing then says stuff
retention_analysis: whatever H3 writer spits out - facial appearance, clothing, reference pictures
detailed_description: whatever H3 writer spits out
overall_soundscape: whatever H3 writer spits out
non_diegetic_music: whatever H3 writer spits out
[chunk 1] character who is <subject 1> does a thing at water location that is <subject 3>
[chunk 2] character with wet hair grabs a thing that is <subject 2>
[chunk 3] character with wet hair holding the thing, then says a thing "a thing"
@gigomikol perfect
it supports hindi?
Yes, LLMs generally understand any language
Why ollama doesn't see igorls/gemma-4-12B-it-heretic-GGUF:latest?
but through Linux console i can run it without problems on ollama
If it runs in Ollama but doesn't appear in Prompt Writer, Ollama may be reporting that GGUF without vision capability. Also, if your Ollama is running inside Linux/WSL, make sure Prompt Writer is connected to that same Ollama host.