Minimax H3 is a very powerful model, but many friends haven't written very good prompts. Now I will share my prompt optimisation t2v workflow. It is based on Qwen 3.6 in llama cpp for prompt optimisation. The optimisation rules are derived from analysing the official H3 user manual through Chat GPT, combined with some of my own understanding. These rules may not be perfect, but in my tests, the video quality generated is quite good. You can also add your own optimisation rules on top of this. Now you can just enter simple prompts like "a little bee collecting nectar in a flower bush" and get quite good videos through optimisation. Of course, you can also write more complex prompts. Finally, I wish everyone a lot of fun and am very grateful that H3, such a great model, is open source.
If your device isn't good enough, you can still try it here:
AI APP: Minimax h3 T2V Qwen3.6 prompt optimisation Experience link: https://www.runninghub.ai/ai-detail/2084834757149773825?inviteCode=d2zvx3lc
Workflow: Minimax h3 T2V Qwen3.6 prompt optimisation Experience link: https://www.runninghub.ai/post/2084826052081537025/?inviteCode=d2zvx3lc
Description
FAQ
Comments (4)
You have in your instructions:
Requirements:
Analyze the character's current action, posture, and body state in the provided image. While maintaining continuity in the character's state, scene, and action logic, naturally extrapolate the character's subsequent movements. Only describe the extended actions; do not repeat actions already depicted in the image.
But in workflow you don't feed the image to qwen?
What you should be seeing is the content in the top box, the content in here will be overridden by the prompt you write and won't work. The rules in the bottom box are the ones that take effect and will optimise your prompt.
This just optimises the prompt you entered, it won't analyse images. So it doesn't connect to images.
Thanks. So I understood correctly it's unnecessary.
Was just wondering whether I got your idea correctly, put your main instruction thing to system prompt to llama.cpp and experimenting without extra nodes in comfy.