MiniMax-H3-Image: 16 MP Text-to-Image and Editing in ComfyUI

ComfyUI Wikinews

MiniMax open-sourced H3-Image: one 33B checkpoint for 16 MP text-to-image and multi-reference editing, 6 steps at CFG 1, native in ComfyUI.

MiniMax-H3-Image is MiniMax's open-weight image model, published on 2026-11-10. One checkpoint generates text-to-image and edits up to 16 reference images, at about 16 megapixels, in 6 steps with no CFG. Native ComfyUI support merged the same day (PR 17655) and ships with official T2I and Edit templates.

Fifteen MiniMax-H3-Image text-to-image results: 3D fantasy battle, landscape painting, tree frog, indigo rabbit, 3D girl, flower portrait, sci-fi city, dumplings, highland cow, mirror portrait, otter and sheep, pixel art cat, jewel beetle, giant cherries, fisheye selfie

Fifteen text-to-image results at 16 MP. Source: MiniMaxAI/MiniMax-H3-Image

What it does

Output size4096x4096 (1:1) up to 6272x2688 (21:9), roughly 16 MP, no upscaler
Sampling6 steps, CFG 1, no negative prompt
Editingsame checkpoint, up to 16 reference images per run
Backbone33.1B DiT (H3 architecture without the audio layers)
Text encoderQwen3-VL-32B, layer 50 hidden states
VAEnew image VAE, 16x downsampling, 24 latent channels

The model expects a long English description of the finished picture. MiniMax's hosted H3-Image-Context-IR rewrites short prompts into that form; offline you can run it with the optional MiniMax-H3-Image-PE prompt expander, or use the 9B Qwen3.5 expander wired into the official ComfyUI template.

Three photorealistic photos: woman hanging sheets on a rooftop at dusk, close profile of a curly-haired man with a cup, flamingos in shallow water

Photo-style output. Source: MiniMaxAI/MiniMax-H3-Image

Edit with reference images

Editing is the same checkpoint: pass images with MiniMaxH3ImageConditioning and refer to them as Picture 1, Picture 2 in the prompt. The 16-image limit covers cases like furnishing an empty room from nine product shots, or turning a painted game map into a first-person view.

Ten reference images of furniture and an empty sunlit room on top, the same room furnished as a living room below

Ten references: nine pieces of furniture plus an empty room. Source: MiniMaxAI/MiniMax-H3-Image

Run it in ComfyUI

  1. Update ComfyUI to v0.45.0 or newer (tagged 2026-11-10). The node is MiniMax H3 Image Conditioning; the T2I and Edit templates are already in the template browser.

  2. Download the Comfy-Org repack from Comfy-Org/MiniMax-H3-Image:

    FileFolderSize
    minimax_h3_image_pruned_int8_convrot.safetensorsdiffusion_models21.0 GB
    minimax_h3_image_pruned_w4a8.safetensorsdiffusion_models12.5 GB
    minimax_h3_image_pruned_bf16.safetensorsdiffusion_models40.2 GB
    qwen3vl_32b_minimax_h3_int8_convrot.safetensorstext_encoders27.1 GB
    qwen3vl_32b_minimax_h3_nvfp4_awq.safetensorstext_encoders15.7 GB
    qwen3.5_9b_w4a8_gptq.safetensorstext_encoders7.3 GB
    minimax_h3_image_vae_fp16.safetensorsvae1.2 GB

    The qwen3.5_9b file is only for prompt expansion through Generate Text, and the Edit template uses it by default.

  3. Load a template and set the sampler to 6 steps, CFG 1. Raise the resolution before you spend a queue: the template starts at 4 MP to stay inside smaller cards, and results below the 4 MP floor get soft.

Both JSONs are the official workflow_templates files, unmodified.

ComfyUI graph of the MiniMax-H3-Image text-to-image template

The T2I template: prompt expander, MiniMax H3 Image Conditioning, 6-step KSampler at CFG 1.

Notes

  • Best around 16 MP. MiniMax puts the working range at 4 to 20 MP; common 16 MP sizes are 4096x4096, 5472x3072 and 6272x2688.
  • Text is a first-class target. Instruction following is the headline claim, and the puzzle-board samples below are the model writing its own layout and labels.
Two logic puzzle boards generated in one pass: a Chinese candidate-selection puzzle and an English CONFIDENTIAL FILE MISSING board with six suspects

Text-heavy boards, generated rather than composed. Source: MiniMaxAI/MiniMax-H3-Image

Resources

Comments

Sign in with GitHub to join the discussion.

Loading comments…
MiniMax-H3-Image: 16 MP Text-to-Image and Editing in ComfyUI | ComfyUI Wiki