Seed hunting workflow for MiniMax-H3-Image. Hunt seeds at 4 MP, pick one, render it at 16 MP. Native ComfyUI nodes only, no custom packs.
Why a hunt at 4 MP and not "same seed, bigger size"
In H3-Image a seed does not keep its picture when the size changes. Seed 42 at 4 MP and seed 42 at 16 MP are two different photos.
So the final stage takes your pick as <Picture 1> plus your full prompt and renders it at 16 MP. Composition, people and text carry over, the 16 MP pass adds the detail.
The full prompt also fixes small things the pick got wrong, as long as the prompt describes them.
The medium picker
Without a stated medium, H3-Image goes for a crunchy, over-sharpened HDR look, worst on faces. MiniMax's own demo Space does the same with a bare prompt.
The dropdown puts a tested medium sentence in front of your scene: 35mm film, full-frame digital, phone selfie, phone main camera, or none.
Film gives the most natural skin. The phone options make convincing UGC stills. Pick none for illustrations, posters and studio product shots and name the medium yourself.
In the gallery: same seed, same scene, with and without the medium, 100% crop.
How to use
Pick a medium, write the scene. Visible text in double quotes. No skin or sharpness words (pores, 8k, hyper-detailed).
Pick the aspect ratio. Final is 16 MP, the hunt runs at 4 MP.
Set the number next to Run to 4 and press Run. Seed is on increment, so you get 4 seeds, saved as hunt_s<seed>.
Type the seed you like into seed, set it to fixed, click the 4 · FINAL group title, Ctrl+M to switch it on, Run once.
Next hunt: Ctrl+M on 4 · FINAL to switch it off, seed back to increment.
The model renders one image at a time (batch size 1), so the hunt is a queue of single runs, not a batch.
Speed, RTX PRO 4500 (32 GB) with 57 GB system RAM: about 6.5 s per 4 MP hunt seed, about 75 s per 16 MP final. ComfyUI streams the 21 GB model from system RAM, so a 32 GB card runs the full int8 model.
Mac: one ComfyUI process can't hold the 21 GB model and the 25 GB text encoder on a 48 GB Mac. Untested on 64 GB and up.
Models, from Hugging Face Comfy-Org
diffusion_models/minimax_h3_image_pruned_int8_convrot.safetensors (21 GB)
text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors (25 GB)
vae/minimax_h3_image_vae_fp16.safetensors (1.2 GB)
Needs ComfyUI master from 2026-11-12 or newer.
Known misses
Small logo text on clothing garbles. Big signs, menus, banners, labels and headlines hold.
Rows of small book spines come out partly gibberish.
Tiny background faces in a crowd stay soft, even at 16 MP.
A label you call "plain" comes out blank. Write the label text.
Text glitches in your pick carry into the final. Pick the seed with the cleanest text.
Gallery: scenes straight out of this workflow, plus a city set rendered with the same model on a Mac (its post notes say what was retouched). Prompts, aspect ratios, the 4 hunted seeds and the pick for each are in demos.json.
Legacy v1/v2 (edit + inpaint)
v1 and v2 used the H3 video model as a single-frame image editor: edit, mask inpaint and a 16 MP route, for Mac and NVIDIA, with custom nodes. They stay downloadable on their version tabs, but their GitHub repo is no longer public and the Mac recipe relied on a turbo LoRA and GGUF files that are now outdated. For new work use v3.0.
Description
v3.0: rebuilt for MiniMax-H3-Image (the dedicated image model) on the official ComfyUI template, core nodes only. Seed hunt at 4 MP, re-render the pick at 16 MP with the pick as <Picture 1>, plus a medium picker that stops the crunchy HDR look on faces. 10-scene demo gallery rendered straight from the workflow. Verified on an RTX PRO 4500 (2026-11-16): ~6.5 s per hunt seed, ~75 s per 16 MP final.












