MiniMax H3 Image day-one tips: one checkpoint for t2i and edits

ggbro_itsover · reddit · 2026-11-11

A hands-on day-one rundown of MiniMax H3 Image (with Day 0 support on Fuobe):

Same checkpoint handles text-to-image and edits: attach reference images to the conditioning node to edit, leave them off to generate.
Plug in the VAE when editing, otherwise references only go through the text encoder and faces/rooms drift more.
References are referenced as <Picture 1>, <Picture 2> etc., in connection order.
Recommended settings: 6 steps, cfg 1, no negative prompt.
It wants long prompts — describing light, lens, and subject placement markedly improves results.
Solid lettering (exact poster text) and native 4096×4096 resolution.

If you already run H3 video, the text encoder carries over — just add the model + VAE.

More from Multimodal