MiniMax H3 is a video model, but if you ask it for a single frame it turns out to be a really good image editor. Give it a photo plus the thing you want in it (a logo, a label, a poster, a mural) and it puts it in properly: wrapped around a can, painted into brick, glowing on a wall with the reflection in the wet street. It also does plain text-to-image, and the 16 MP workflows take either one up to about 16 MP while staying sharp when you zoom in.

The trick comes from Patient_Ratio4177 on r/StableDiffusion. I packaged it into ComfyUI workflows for Mac and for NVIDIA.

What you get

Six workflows:

• Edit (Mac) and Edit (CUDA): two images in, one edited image out

• Inpaint (Mac) and Inpaint (CUDA): paint over the part you want changed. Everything outside it stays untouched, apart from a soft blend at the edge.

• Edit 16 MP (Mac) and Edit 16 MP (CUDA): the edit workflow at 4 MP, then a 2x latent upscale and a 16-step refine, so about 16 MP comes out. Asking H3 for 16 MP in one go looks fine from a distance but melts when you zoom in; this route stays sharp.

Each one opens with an example already loaded (the demo images are in the zip). The inputs you actually touch are grouped at the top left, and there's a how-to note right there.

Getting it running

1. Copy the images from the zip's input/ folder into ComfyUI's input/ folder

2. Install the small h3_single_frame node from the GitHub repo and restart ComfyUI. It's what lets H3 render one frame instead of a video.

3. Mac: you also need ComfyUI-GGUF, ComfyUI-ClipProj and comfyui-obvpm. For inpainting (Mac or CUDA): ComfyUI-MAINodes. For the 16 MP workflows: Comfyui_Minimax_h3_latent_Upscaler (LBH-123-AI) and its 0.7 GB model.

4. Grab the models. The repo README lists every file, with links and where it goes.

Speed: about 11 minutes for a 4 MP edit on an M5 Mac, about 5 minutes on an RTX PRO 4500. The 16 MP workflow takes about 10 minutes on the RTX PRO 4500. CUDA needs a Blackwell card (RTX 50xx / RTX PRO) on driver 580 or newer.

Writing prompts that work

• Start with Task: Reference-guided generation.

• Refer to your images as <Picture 1>, <Picture 2> (first loaded image = Picture 1)

• Say what each picture is for, and what it is NOT for: "<Picture 2> is the label artwork only. It does not supply a background or lighting."

• Say "exactly one" when you want one of something

• List what has to stay the same in Picture 1

• Go to 4 MP if there's small lettering. At 2 MP small text turns to mush.

What it's not good at

• A normal edit redraws the whole picture, so fine details can shift a little. If the rest has to stay pixel-perfect, use the inpaint workflow.

• Tattoos look like stickers (last image). Detailed illustrations get distorted, but text and logos hold up.

• Inpainting can leave a very faint grid on smooth areas like sky or plain walls

• Faces from a reference photo don't hold a real likeness when they're small in the frame

• It's not an upscaler for existing photos. Qwen-Image-Edit 2.1 does that job well.

Want to script it?

The repo also has a command-line tool that runs the same workflows: batch edits, several seeds at once with a contact sheet, extending an image past its edges, automatic face fixes, and building huge images one region at a time. pip install git+https://github.com/Bambushu/h3image

Repo, full docs and model list: github.com/Bambushu/h3image

All brands in the demos are made up and every demo image is AI-generated.