Simple MiniMax H3-Image workflow (text to image)
A few people asked about blurry or weird eyes with the official ComfyUI template. The template has two problems: it renders at only 4 MP, and it uses euler. H3-Image works with big 32×32 px patches, so at 4 MP the eyes in a full-body shot are only a few patches wide and turn to mush. Euler also gives noisier skin and makes faces look older.
This is a stripped-down workflow that fixes both:
16 MP output (a half-size copy is saved for viewing)
SamplerLCM (s_noise 1.10 → 0.90), simple, 6 steps, cfg 1
decodes with the P-H3 VAE for a bit more detail
Only core nodes, ComfyUI 0.45 or newer.
The download links are in the note inside the workflow. If you want to use a LoRA, there’s an optional slot: pick your LoRA, select the node and press Ctrl+B. The same goes for the full 16 MP save. You can also drag the example image into ComfyUI to get the workflow with its exact settings.
Reference Image Ready!
Add a Load Image node and plug it into image_0 on the conditioning node. That’s it, the VAE is already connected. In your prompt the images are <Picture 1>, <Picture 2> and so on. Picture 1 is the one that gets edited, or you can just use it as a reference for a new scene. There’s a short note in the workflow with examples.