⚙️ MiniMax H3 Prompt Enhancer — V7.3 Turn simple ideas into detailed, structured prompts for MiniMax video generation.
This ComfyUI workflow is a prompt enhancer built around the official MiniMax H3 prompting guide. It helps transform short, rough ideas into clearer and more effective video prompts — without any complex LLM installation.
Supports T2VA, I2VA, FL2VA, L2VA, and REF2V workflows.
List of models: https://huggingface.co/DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF
Choose the best for your hardware. *MTP models incompatible with the Custom LLM node*
Custom nodes needed (also in comfyui manager):
LLM Text Processor: ⚠️*Must be insalled manually from here, there is a bug for that node in the node manager*⚠️ https://github.com/KingManiya/ComfyUI-LLM-text-processor
Comfyui-RMBG: https://github.com/1038lab/ComfyUI-RMBG
rgthree-comfy: https://github.com/rgthree/rgthree-comfy
✨ What's New in V7
Added a video input node and a second QwenVL node (automatic download on first run for this model) with instructions on how to use.
Backed prompt refining.
Model Change. Now using Qwen3.5-9B-The-Defiant-Fable-Uncnr-Heretic-NEO-MAX-Q8_0. (Download directly in workflow. Extremely accurate, runs fast and easy on 16 VRAM)
V7.1
Separated REF2V Prompt in backend for better accuracy.
Removed time stamps instructions for video reference as it was exceding the video duration most of the time.
V7.2
Prompt cleaning.
V7.3
Prompt cleaning.
Added Number of shots and video duration nodes, no need to input manually now!
🎯 The Goal
This workflow isn’t meant to blindly generate the “perfect” prompt for you. Instead, it gives you a well-structured starting point that you can quickly review, tweak, and customize before sending it to MiniMax.
Simple idea → Enhanced prompt → Your edits → Video generation
💡 Tip: The generated prompt is a starting point. Read through it and adjust the details, actions, camera movement, timing, and style to match your exact vision.
Feel free to use, modify, and integrate this workflow into your own projects. If you create something with it or improve the prompting method, share your results and prompting tips in the comments so the community can keep improving it!
If you use this workflow in your own workflow pack, a quick mention would be greatly appreciated. ❤️
Description
Changed some backend nodes to further reduce custom node dependencies.
Added better instructions for REF2V and why it does not need a video/audio input to work as intended.
Added the model choice to front end of the workflow.
FAQ
Comments (25)
Thanks for the helpful hints.
This is the right solution.
I used chatgp and others before, and they wrote long hints, and the results weren't too bad.
But your hints are short and precise.
This will really help people!!!
Thank you for the sweet comment. Glad it helps!
TA...YES, IT HELPS. THANKS!
........BUT HERE'S YOUR NODE...
Image Compare (RMBG) 🖼️🖼️-THEY CUT IT FROM 3 TO 2)
WILL HAVE TO CONNECT = LOOK FOR ANOTHER ONE
well my good sir. thanks for 6.1 works great now all i prompt is better translated to H3 language. i "fixed" the no video think . i feed in picture 2 a some frames of the video and i get a good result .
"""
describe the woman from character sheet.
the woman from character sheet will replace a woman in video1. you don't have video1, you have 1 image from video 1 which is picture 2.
you dont mention picture 2 in the prompt.
motion movement and camera will be the same as video1.
nothing will change from video1 except the woman in the video. the video must reamain the same with background and camera movement and motion and movement except for the woman.
Do not change anything else in Video 1.
"""
i think if you added a node to give like 4 frames or more from a video in diferents area of the video or decide how many frames to give the llm you can bypass the llm not being able to take video in this version. may be other ways just saying how i was able to "fix" my issue. thanks a lot and regards
I think it is pretty cool, thanks for the simplifications, much easier to use it now and works flawless (still testing latest 6.1 tough).
One more hint for next version: comfy-image-saver seed node also can be replaced with a native one: Comfy - utilities - seed.
One less dep. Thank you for your work!
will be don e in the next verison thx :)
@KiraNugget Thank you for sharing your work sir. And as @tsgeza870 said, nothing best than native nodes, in a more general (absolutely not personal) critic it's always "...sigh..., is this custom node really required? where is it linked/called ? will this custom repo install some unwanted dependencies or create duplicates, etc...doesn't this native node do the same ?....".
I get error
# ComfyUI Error Report ## Error Details - Node ID: 39:6 - Node Type: LLMTextProcessor - Exception Type: RuntimeError - Exception Message: RuntimeError: llama.cpp inference failed with exit code 1: ggml_cuda_init: failed to initialize CUDA: (null) load_backend: loaded CUDA backend from C: ....
## System Information - ComfyUI Version: 0.30.2 - Arguments: ComfyUI\main.py --lowvram --windows-standalone-build - OS: win32 - Python Version: 3.12.10 (tags/v3.12.10:0cc8128, Apr 8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)] - Embedded Python: true - PyTorch Version: 2.8.0+cu128 ## Devices
Seems like your comfyui and pytorch are not up to date
@KiraNugget what version of pytorch is needed?
@metalwolfeagle812 im at work right now i cant tell what version i have but i know its with cu130. Best bet is to go for a clean instal with comfyui-easy-install app. It updates everything in one click. If you already have the app you can change version somewhere in one of the folders inside the main easyinstall folder
may i suggest you something ? well i'll do it anyway, instead of a 27b 2bit, you should tell people to go with Qwen3.5-9B-The-Defiant-Fable-Uncnr-Heretic-NEO-MAX-Q8_0, same weight as the 2bit 27b and way better.
Ill try it, thanks!
@KiraNugget maybe i weirded it a bit weirdly, so what i might you can have the full Q8 Qwen3.5-9B which is super good versus only a 2bit of a 27b ! of course if you run the 27b Q8 that will be better but.. yeah.. gonna have a beast. so.. Qwen3.5-9B Q8 is a big winner :).
Qwen3.8-27B-Heretic-Q4_K_M.gguf?
@vicaut can you make a sentence ? i've no idea what you're asking ._.
@EvilFear I'm not asking - use my suggested model.
@vicaut hard to tell with a single name and a question mark, 'communication' you know that word ._. ? anyway you're missing my whole point, a liter and faster model that is more pc budget friendly, and better run a 9b Q8 than a 27b 2Q. your "suggestion" is an even heavier model and thus denied. on that note, i do not wish to continue that conversation ^-^/
I guess Qwen3.8-27B-Heretic-Q4_K_M.gguf is now better than the default choice in this workflow.
Downloading it right now to test, thanks for the hint!
Q4_K_M is good but slow for my hardware at 16vram. With complex promt for ref2v it timed out after 300 seconds. I tried Q3 but it only outputed his thinking process even when explicitly told not to.
@KiraNugget True, but for a 4090, 5080 or 5090 it is the best option.
@vicaut yes, But as a base model, I think most people do not own high end gpus like this yet. So yeah I will probably put it as a second model in the workflows download list for the rich boys like you! ;) Just kidding, model is good but most people won't be able to use it and will scream in the comment that the model is too big lol
i am getting this when i run it on a H100 server gpu,
"# ComfyUI Error Report
## Error Details
- Node ID: 39:6
- Node Type: LLMTextProcessor
- Exception Type: RuntimeError
- Exception Message: RuntimeError: Automatic llama.cpp binary download currently supports Windows x64 CUDA 13 only. Other platforms are intentionally isolated behind the platform mapping for future support." is node not compatible?
you need to manually install the node from the link in the description, and you need to update to cuda 130