MiniMax H3 is not limited to 9 reference images.
I have not yet found the actual upper limit, but I suspect it may be as high as 256. The 9-image limit appears to come from a schema inside ComfyUI rather than from H3 itself.
This workflow includes a .bat file that patches ComfyUI and removes that 9-reference restriction, allowing you to experiment with much larger reference sets.
So far, I have tested it successfully with 15 reference images at once.
Stop wasting time trying to make H3 understand a character sheet (been there).
Now simply reference all the views, you have plenty of references
Description
FAQ
Comments (6)
WHAT IS THIS WITCHCRAFT!
I actually have a legitimate question. Do we know how reference images impact VRAM and RAM usage? It would seem that the more reference images you use, the better system you need.
You can scale down the references to fit more if you need to. I usually scale my stuff down to 0.81 megapixels for no particular reason. Usually set video references to 0.5 megapixels. Of course, the higher the megapixels for the reference, the better quality you might get. Just depends on what you're trying to do. You can pick and choose the size of each input.
I don't think it matters as much as the length and resolution of the clip. RefMods https://comfyui-wiki.com/en/news/2026-09-07-minimax-h3-refmods showed that once inferred 256 images only take up a few megabytes.
I was able to generate this test example using a 5090
Has anyone trained with success more than 9 images with 16 gb vram?
It is not really training but referencing. I don't think the impact of adding more reference images is much higher, check this other project:
https://comfyui-wiki.com/en/news/2026-09-07-minimax-h3-refmods
Problem is binding the audio as audio swapping is a major issue, can you look into that?