Photographic edits with MiniMax H3 Image
Two ComfyUI workflows for MiniMax H3 Image edits that look like photographs. They keep the face of the person in your picture, and give real skin, stray hairs and fabric weave at 16 MP.
H3 Image - realism edit uses the H3 Image KSampler node (comfyui-h3-sampler, GPL-3.0): one node with tested defaults.
H3 Image - realism edit (RES4LYF) builds the same recipe from stock nodes and RES4LYF, with every setting in view.
For a given seed both give the same composition and the same look, with small differences in fine detail. The examples were picked by eye, and each carries its workflow: drag it into ComfyUI to open it. The Limits section lists what failed.
New: Text to image v1.0, the same recipe for a prompt alone (pick it at the top of the page). See Text to image below.
Why it looks real
H3 Image is step-distilled on a fixed 6-step noise grid. Composition, objects and identity are set above noise level 0.5. Photographic texture appears only at 16-20 MP, so the workflow renders at 16 MP.
Structure: plain Euler on that grid, the way the model was distilled. In a 10-seed test, other samplers there drew a duplicated object (a second umbrella shaft), and fresh noise there cost likeness.
Texture, on the last two steps: a second-order exponential step (res_2s_rkmk2e) with fresh noise (eta 0.65) and a 7% lower noise level on its second evaluation. This brings out pores, stray hairs and grain.
References: a tight crop of the face as Picture 2 was the biggest likeness gain we measured.
Setup
A ComfyUI recent enough to have the MiniMax-H3-Image templates (tested on 0.45.3).
Models, which ComfyUI offers to download when missing: the diffusion model (21 GB), the int8 text encoder (27 GB) or, on Blackwell GPUs, the NVFP4 one (15.7 GB), and the VAE (1.2 GB). The README in the download links each file.
Node version: git clone https://github.com/igorls/comfyui-h3-sampler inside ComfyUI/custom_nodes, then restart ComfyUI. RES4LYF version: install RES4LYF from ComfyUI-Manager.
To run the example as it is, copy the pictures from the download's input folder into ComfyUI/input.
Use
Picture 1: the photo to edit.
Picture 2: a tight crop of the face in Picture 1, with some hair.
Prompt: keep the first three sentences and the last three. Replace the scene in between: framing, pose, action, expression, clothing, place, light.
Pick the aspect ratio, leave 16 MP, and Run: about 30-50 s per image on an RTX PRO 6000 (96 GB).
The notes on the canvas also cover new text in the picture, products, and extra reference pictures.
Limits
Very strong expressions (shouting, crying) move the face the most.
Text on signs and packaging came out right in our tests. Printed text on clothing is the weak spot: it can look pasted on, or come out ghosted or misspelled, depending on the seed, so check it at 100%. Stock KSampler draws the same thing, so it comes from the model, not the sampler.
Tested only on a 96 GB RTX PRO 6000 (Blackwell).
Text to image (new)
The Text to image v1.0 version on this page makes 16 MP photographs from a prompt, with the same models and sampler as the edit workflows. Keep the prompt's first sentence ("One ultra photorealistic photograph.") and its last ("Sharp skin detail, no text, no watermark."), and describe the picture in between: who or what, the action and expression, clothing, place and light.
It keeps H3 out of its crusty text-only mode. With no picture in its conditioning, H3 Image renders in a text-only mode that comes out crusty and over-sharpened at 100%. Any picture switches it to its far cleaner reference mode. The workflows use Kijai's trick from KJNodes (the "text with empty image" mode of MiniMax H3 Prompt Weight): a black picture's vision tokens in front of the prompt on the first two steps only. The node version uses H3 Image Empty Picture, our port of it in comfyui-h3-sampler 1.1, pixel-identical to Kijai's node (if you already have the pack, run
git pullin its folder). The RES4LYF version builds it from stock nodes. With the same prompt and seed, fine-detail energy fell from 35 to 12 on a landscape and from 16 to 9 on a portrait, the range of our photo edits.About 27 s per 16 MP image on an RTX PRO 6000 (96 GB), once the models are loaded.
Limits: hands, held objects and text can come out doubled, ghosted or malformed, depending on the seed, so render a few and check at 100%. Props are the weak spot: every reject while we built the examples was a prop (surfboard fins on the deck side in 3 of 3 seeds, patterned bowls ringed with made-up characters, a blade-like jug spout, an odd canoe plank). None for skin or texture.
A lab
This page doubles as a lab: we will post the same shots with other samplers, settings and styles here, each image with its workflow embedded.
The people in the examples were generated with Krea 2, the brands are invented, and the anime character is an original design made with H3 for this page.
Credits: MiniMax H3 Image by MiniMax, ComfyUI, and RES4LYF by ClownsharkBatwing.
Lab 1: how big can H3 Image go?
One 96 GB RTX PRO 6000 (Blackwell), this workflow's sampler defaults, seed 5000, and two edits: a waist-up portrait and a wide shot. Sizes run from 4 MP up to 179 MP (10912x16384, the most the H3 conditioning node accepts).
Memory is not the limit. Peak use was 38 GB at 4 MP, 47 GB at 96 MP and 55 GB at 179 MP, and about 37 GB of that is the model files. ComfyUI never had to decode in tiles or offload the model.
Time: 28 s at 16 MP, 54 s at 32 MP, 2.2 min at 64 MP, 4.1 min at 96 MP, 6.6 min at 128 MP and 11.6 min at 179 MP.
Quality holds to 96 MP. At 100%, 64 and 96 MP show single hairs, pores and fabric weave. From 128 MP, fine textures (stripes, hair, knit, steam) break into blocks the size of a model token, and at 179 MP they appear everywhere. The picture as a whole stays coherent at every size.
It is the model, not the setup. At 128 MP we got the same blocks every way we tried: decoding the same latent in small windows, ordinary PyTorch attention instead of the int8 attention kernel, and the bf16 checkpoint instead of the int8 one. Our reading is that H3 maps every image to the same coordinate range. At 128 MP neighbouring tokens sit about 5.5 times closer together than at its native 4 MP, and the model can no longer keep them apart.
Practical range: 16-64 MP for everyday work. 96 MP is the ceiling, and more VRAM does not raise it. The workflow's size node stops at 16 MP; to go bigger, disconnect it and type width and height on the Prompt node (multiples of 32).

