MiniMax-H3-Image in ComfyUI: 16 MP Text to Image and Editing (Files, VRAM, Settings)

MiniMax has released MiniMax-H3-Image, an open-weights image model that makes pictures at about 16 megapixels, which is 4096 x 4096 for a square image. The same checkpoint also edits images, using up to 16 reference pictures.

ComfyUI supports it natively from v0.45.0, released on 10 November 2026, with two ready-made templates: text to image, and image edit.

This guide covers which files to download and where they go, which precision fits your GPU, the exact settings, how editing with <Picture 1> tags works, prompting tips, the license, and fixes for the errors people hit first.

What MiniMax-H3-Image Is

MiniMax-H3-Image is the image member of the MiniMax H3 family. According to the ComfyUI Wiki model page, it is a 33.1B dense single-stream diffusion transformer, described as the H3 model without its audio layers.

Key facts from the official sources:

MiniMax says a 4K image takes “about 10 seconds” on its own serving setup. That figure comes from MiniMax, on data-centre GPUs, and it is not a number you should expect on a home card. MiniMax has not published benchmark scores.

H3-Image vs P-H3-Image vs H3 Video

The names are close, so here is how they differ:

Because H3-Image shares the Qwen3-VL text encoder with H3 video, you can reuse that file if you already run H3.

Model Files and Folders

Comfy-Org repackaged the weights for ComfyUI on Hugging Face. You need one diffusion model, one text encoder, and the VAE. The prompt expander is optional.

Diffusion model (pick one)

From Comfy-Org/MiniMax-H3-Image, into ComfyUI/models/diffusion_models/:

Text encoder (pick one)

From Comfy-Org/MiniMax-H3 (the video repo), into ComfyUI/models/text_encoders/:

VAE

minimax_h3_image_vae_fp16.safetensors (1.17 GB) into ComfyUI/models/vae/.

Prompt expander (optional)

The official templates use qwen3.5_9b_w4a8_gptq.safetensors (7.34 GB) from Comfy-Org/Qwen3.5, into ComfyUI/models/text_encoders/. MiniMax’s own expander, MiniMax-H3-Image-PE, is 17.9 GB in NVFP4, which is why Comfy chose the lighter 9B model.

Your folder should look like this:

ComfyUI/models/
  diffusion_models/
    minimax_h3_image_pruned_int8_convrot.safetensors
  text_encoders/
    qwen3vl_32b_minimax_h3_int8_convrot.safetensors
    qwen3.5_9b_w4a8_gptq.safetensors      (optional expander)
  vae/
    minimax_h3_image_vae_fp16.safetensors

Our ComfyUI model folders cheat sheet explains every folder if you are new to this.

Which Precision Fits Your GPU

Neither MiniMax nor Comfy has published VRAM figures for ComfyUI yet, so treat this as a guide based on file sizes, not test results.

ComfyUI runs the text encoder first, then moves it to system RAM before sampling, so the diffusion model is what has to fit on the card. Large images also need room for activations.

System RAM matters a lot here. The int8 text encoder alone is 27 GB, so with 32 GB of RAM your PC may swap to disk. 64 GB is a sensible minimum and 96 to 128 GB is comfortable.

The AI/TLDR release summary says SGLang offloading recipes go down to 12 GB cards. That is a different runtime, but it shows that heavy offloading can work on small GPUs at the cost of speed. To check how much memory a run actually uses, see how to check GPU VRAM usage.

Step-by-Step Setup

1. Update ComfyUI to v0.45.0 or newer. In ComfyUI Manager use Update ComfyUI, or run git pull in the ComfyUI folder and pip install -r requirements.txt. Our safe update guide covers Desktop, Portable and Git installs.
2. Update the templates to 0.11.104 or newer: pip install -U comfyui-workflow-templates inside the ComfyUI Python environment.
3. Download the files from the tables above and place them in the right folders.
4. Restart ComfyUI and refresh the browser.
5. Open a template: Workflow, Browse Templates, then search “MiniMax-H3-Image”. Pick Text to Image or Image Edit.
6. Point the loaders at your files. If you chose w4a8 or nvfp4, select those names in the Load Diffusion Model and Load CLIP nodes, because the template defaults to int8.
7. Run. The first run is slow while the files load from disk.

The templates are also on GitHub: text to image and image edit. Drag the JSON onto the canvas to load it.

Text to Image Settings

From the official ComfyUI tutorial:

The steps and guidance are baked into the distilled checkpoint. Changing them does not give better images. MiniMax’s API even rejects other values.

Resolutions

Every edge must be a multiple of 32 pixels. The native 16 MP sizes are:

The template opens at 4 MP so smaller cards do not run out of memory. Its resolution selector sets the aspect ratio and the megapixels separately, so you can raise the size step by step. If width and height are left at 0, the conditioning node uses 2048 x 2048.

Below about 4 MP, MiniMax’s reference code generates on a larger canvas and resizes down, because quality drops at small sizes. So do not expect sharper results from 1 MP than from a dedicated small model.

Editing With Reference Images

The same checkpoint edits images. There is no separate edit model or adapter.

• The image you want to change is <Picture 1>.
• Extra references are <Picture 2> to <Picture 16>, numbered in the order you connect them.
• Refer to them by label in the prompt, for example: “Replace the car in <Picture 1> with the bicycle from <Picture 2>, keep the street and lighting.”

To add a reference, add another Load Image node and connect it to both the prompt generation step and the MiniMax H3 Image Conditioning node.

The VAE must be connected to the conditioning node when editing. Without it, the references only reach the text encoder and the result will not keep the look of your source.

You do not need to resize inputs. The template rounds the source to a multiple of 32. Internally ComfyUI uses a 1024 px short edge for the vision side and clamps the VAE side between 1024 and 2048 px.

If you run out of memory while editing, enable the template’s Resize Image/Mask node and lower its megapixels.

Good edit jobs for this model: product swaps, changing materials or colours, replacing signs and text, and combining a character from one image with a scene from another. For masked edits, which H3-Image does not support yet, see our ComfyUI inpainting guide.

Prompting Tips and the Prompt Expander

H3-Image was trained on long, detailed English captions. Short prompts work, but detailed ones follow your intent better.

• Describe the whole picture: subject, setting, lighting, camera angle, lens feel and mood.
• Put text in quotes if you want it rendered, such as a sign reading “OPEN LATE”.
• Say what you do not want in the positive prompt, such as “clean background, no text”, because there is no negative prompt.
• Use the expander for short ideas. In text to image it is off by default. Turn it on and it writes a long English caption and picks the aspect ratio. In the edit template it is on by default and writes a description of the finished image that mentions each reference.

Example short prompt: “A cosy bookshop at night, rain on the window, warm lamps, a cat asleep on a stack of books, a sign reading ‘OPEN LATE’.” The expander turns that into a paragraph that names the lens, light colours and layout.

License and Commercial Use

The weights use the MiniMax-H3-Image Community License Agreement, in the repo’s LICENSE file. It is not a standard open-source license like Apache 2.0.

The GitHub page says Section IV requires separate written authorization, requested through platform.minimax.io/license. Read the license yourself before using images commercially, and keep a copy of the version you agreed to.

Current Limitations

• No mask inpainting and no transparent backgrounds yet, per the GitHub README.
• No benchmark scores from MiniMax, so its claims of the best instruction following among open models cannot be checked yet.
• Heavy downloads: even the smallest set is about 30 GB.
• No GGUF yet. If community GGUF files appear, they will need a GGUF loader. Our ComfyUI-GGUF guide explains how those work.

Troubleshooting

“MiniMax H3 Image Conditioning” node is missing

Your ComfyUI is older than v0.45.0. Update it, then restart. On Comfy Cloud, the nodes arrive after the stable release.

I cannot find the template

Update the templates package to 0.11.104 or newer with pip install -U comfyui-workflow-templates, or download the JSON from GitHub.

Out of memory (CUDA OOM)

Lower the megapixels first, then switch to the w4a8 diffusion model. Close other GPU apps and make sure you have plenty of system RAM. For editing, enable the Resize Image/Mask node.

The first run takes minutes

That is normal. ComfyUI is loading 30 GB or more from disk. Later runs reuse the cached models. A fast NVMe drive helps a lot.

Black or grey image

Check that you loaded minimax_h3_image_vae_fp16.safetensors and not an H3 video VAE or another model’s VAE. A wrong VAE gives blank output.

Edits ignore my reference image

Connect the VAE to the conditioning node, and make sure your prompt names the reference by its label, such as <Picture 2>. The numbering follows the input order.

Error loading the nvfp4 or w4a8 file

These low-bit formats depend on GPU support. If they fail on your card, use the int8 files, which the official tutorial uses.

Washed-out images after changing CFG or steps

Put them back to CFG 1 and 6 steps. The model is distilled for exactly those values.

FAQ

Is MiniMax-H3-Image free?

The weights are free to download and run locally under MiniMax’s community license. Commercial use may need written permission under Section IV. MiniMax also sells API access.

Can it run on a 16 GB GPU?

Probably, with the 12.5 GB w4a8 model, a 2 to 4 MP canvas and a lot of system RAM, but there are no official 16 GB tests yet. A 24 GB card with the int8 model is the safer starting point.

Do I need a negative prompt?

No. The model does not support one. Describe what you want to avoid inside the positive prompt.

Can I reuse my MiniMax H3 video text encoder?

Yes. H3-Image uses the same Qwen3-VL-32B text encoder files from the Comfy-Org/MiniMax-H3 repo.

How many reference images can I use?

Up to 16 in total, including the image being edited.

Can I train LoRAs for it?

MiniMax has not released official LoRA training code for H3-Image yet. Watch trainers like ZenTrainer, which already support H3 video.

Is it better than Qwen-Image?

It renders at a much higher native resolution, but there are no shared benchmarks yet. For lower-VRAM setups, Qwen-Image 2.1 GGUF is still the easier choice.

Sources

• ComfyUI docs: MiniMax-H3-Image tutorial (files, settings, templates)
• Comfy blog: MiniMax H3 Image is now available (10 November 2026)
• MiniMax-AI/MiniMax-H3-Image on GitHub (license, sizes, limits)
• Comfy-Org/MiniMax-H3-Image on Hugging Face (repackaged files)
• ComfyUI Wiki: MiniMax H3 Image (architecture and node names)
• AI/TLDR: MiniMax H3-Image release summary (speed claims, hardware notes)