⚠️If you encounter a bug before commenting, try to redownload the workflow first as I am actively fixing bugs as they are reported. If it doesnt work let me know and I'll try to help.⚠️
⚙️ MiniMax H3 Prompt Enhancer — V7 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)
🎯 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
Better instruction prompt.
Added REF2V Prompter.
FAQ
Comments (42)
Works great, thanks!
Thanks to you for the tip! :D
I couldn't get this to work :(
The thing says "prompt executed" and all that, but no text shows up on the right side.
maybe try a different preview text node. I used the one from comfyui-utils-node pack but any preview or show text should work. If that doesnt work, maybe your comfyui and/or nodes needs an update.
there is a bug with gen text in comfi latest version see here
https://github.com/Comfy-Org/ComfyUI/issues/15143
@KiraNugget Ah yea, looks like your workflow is just fine, it's a known bug in comfy itself aaaaa. Ah well, once they get stuff sorted out on their side i'll do an update here :D
@dirkzen it works fine on my side and most people it seems and everything is up to date. Maybe a fresh comfyui install could help.
It works but I'm not really getting much better prompts
It is more about getting the right format of prompt for MMH3 since it can be a bit complex otherwise. I recommend editing the prompt description and action to your liking afterwards.
Maybe your prompt is already very good, idk :/
my codex sub and wallet are empty so i had to get this, its amazing! thanks
Glad you like it!
Thanks A LOT! the conversion is near perfect for me and I can really make my videorenders pop more now :)
For some reason i cant get it to work properly, no matter what model i use, qwen 3.5 4b produces repetitive tags (4k,4k,4k,4k x1000), qwen 32b produces only literal commas ( ,,,,,,, ), what model am i supposed to use?
I have the Same issue. IT doesn't Matter what model I choose... only literal commas.
I use qwen 3.5 4b. I remember having this happened to me months ago with qwen, I dont remember what fixed it though... Sorry I can't help :(
@KiraNugget , tThank you for your advice regarding the model.
I've tweaked the workflow a bit. I switched the prompt generator to Ollama and changed the text preview field—now it's working perfectly for me.
Damn, this happened to me as well at some point. Still can't figure out what's wrong.
It is useful, in my experience you still need to clean the prompt but it is a good starting point.
It probably needs a larger model to do a better job.
For FL2VA, do you just pass a batch of 2 images to the "image" input? I am trying that but not sure it is working.
It can only take 1 image, 1video and 1 audio so in the prompt I instructed it to not give details about references it dosn't have and only output description given to him
Any recomendation for a good model that runs well on 16gb vram?
@libenlubenschlagen Thanks I will try that!
@libenlubenschlagen I feel a bit dumb but I can't find a node to run this model since all the genereate text nod I found/have uses a clip a not a gguf.
You have to load the model as a llamacpp server. See https://www.glukhov.org/llm-hosting/llama-cpp/
Then you use a openai compatible node to connect to the server. eg. https://github.com/hekmon/comfyui-openai-api
Use a client node and give it you base url. for me thats http://127.0.0.1:8070/v1 because i use port 8070. Then connect it to a text completion node and fill out modelname, systemprompt and prompt. Now if you run the workflow it will connect to the running lamacpp server and execute whatever prompt you gave it.
i also use llama-swap because i run multiple model but its not strictly necessary: https://github.com/mostlygeek/llama-swap
Btw if you want the model to be vision capable you need to load the visionprojektor alongside the gguf file into llamacpp
https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF/blob/main/mmproj-F16.gguf
@libenlubenschlagen thank you! It seems a bit user unfriendly tho for someone that would just want to download a plug and play workflow. I will probably do a separate workflow for people who don't mind the extra steps.
@libenlubenschlagen I found a way to use the model directly in comfyui, it is extremely fast and accurate, I will include it in the next update! TYSM!
@KiraNugget The only caveat with this model is that answer quality drops off sharply after a certain amount of tokens have been generated but i dont think a videoprompt would ever get that long
@libenlubenschlagen I have an awesome qwen3.6 gguf model coming in verison 5. stay tuned!
@KiraNugget i know since i was the one who recommended it to you lol
@libenlubenschlagen oh! 😅
I tested a lot of models and this is the best one so far for nsfw because it always executes the prompt exactly as written. Pretty much all other "uncensored" models seem to still retain some amount of safety alignment and always soft censor sexual topics by writing things like "she started touching her crotch area" instead of "she started touching her vagina" or some variant of this behaviour.
https://huggingface.co/HauhauCS/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive
I'm using your workflow extensively now, thing is, sometimes it 'glitches' and no new prompts are generated, only after a F5, can you look into that? (changing seeds don't fix it)
It never happened to me so I cant really know whati is going on. And honestly im not super knowledgable when things like that happens in the backend of comfyui. I usually send reports/problems to grok for a quick fix and update everything. Sorry I cant be of much help right now, plus im at work for the next 7 hours.
@KiraNugget It's ok dog! I love your workflow anyway, make my life so easy as I don't have to deal with online LLM's or installing apps just for it, thanks anyway :)
@xDegenerate I recommand checking dasiwa's workflow. He has a node with mmh3 prompt builder integrated, it pairs well with the outputs of my workflow
Wow, this is amazing! This will save a huge amount of time—thanks!
Thanks to you for the tip! Cheers! ❤️
Absolute must have. If you didn't know how to prompt, you'll be amazed. If you did know and spent good amounts of time on a prompt, you'll be amazed too, because this saves a lot of time. Good approach and good setup. kudos!
Thank you! I will also add a section to expand short concepts soon. Stay tuned.
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.