What you'll get
MiniMax H3-Image is available in ComfyUI with official templates. ComfyUI has supported it natively since 0.45.0, with one template for text-to-image and one for image editing that can take up to 16 reference images. This guide helps you figure out which version your graphics card can handle, then covers downloading, file placement, opening the templates, and avoiding out-of-memory errors. It ends with a prompt you can give to your AI agent for checking and downloading.
The most important point up front: check your VRAM before you pick a version. The official version's model files alone run over 50 GB. If you download it and your GPU can't run it, you've lost a day for nothing.
The full path looks like this: check your VRAM and choose a version, update ComfyUI to 0.45.0 or newer, download the four model files into the right folders, open the official template, and start at 4MP before scaling up based on your VRAM.
Step 1: Pick the right version
ComfyUI currently gives you two options:
| Version | Output size | Download size | VRAM | Best for |
|---|---|---|---|---|
| MiniMax H3-Image official template (INT8) | Native, about 16MP | About 55 GB (four files) | The official docs don't list a peak. The template defaults to 4MP to save memory | People with large professional GPUs |
| P-H3-Image (Pruna's slimmed-down version built on H3) | About 8MP, generated in 8 steps | Smallest set is about 20 GB | The official note puts peak usage at about 21 GB at 8MP | Consumer GPUs with 24 GB of VRAM |
P-H3-Image accepts up to 8 reference images for editing, while the official H3-Image accepts 16. Both use MiniMax's community license, so read the terms before using either one commercially.
Some people in the community report running it on 6 GB or 8 GB cards. They do it by moving parts of the model in and out of VRAM in batches. It's slow, the official docs don't cover that setup, and we haven't tested it ourselves.
Step 2: Update ComfyUI
The official templates require ComfyUI 0.45.0 or newer and template pack 0.11.104 or newer. If you can't find the two MiniMax-H3-Image templates in the template library, your version is probably too old. Comfy Cloud and the desktop build stick to the stable release, so they may get the update a few days later.
Step 3: Download four files and put them in the right folders
The official template uses these four files:
| File | What it is | Size | Put it in |
|---|---|---|---|
minimax_h3_image_pruned_int8_convrot |
Main image model (INT8) | 21 GB | ComfyUI/models/diffusion_models/ |
qwen3vl_32b_minimax_h3_int8_convrot |
Text encoder | 25.28 GB | ComfyUI/models/text_encoders/ |
qwen3.5_9b_w4a8_gptq |
Prompt expansion model | 7.34 GB | ComfyUI/models/text_encoders/ |
minimax_h3_image_vae_fp16 |
Image decoder (VAE) | 1.17 GB | ComfyUI/models/vae/ |
Every file name above ends in .safetensors.
These files are big and have long names, so putting one in the wrong folder is the most common mistake. Have your AI agent check first, then download:
Sign in to unlock this key section
New accounts get 10 free credits, plus 1 more for each daily check-in. Spend 1 credit to unlock and it stays open for you.
Why this order: checking your GPU and disk before downloading keeps you from pulling down tens of gigabytes only to find out it won't run. Reviewing the file list before anything downloads means a misplaced file gets caught right away. And skipping third-party nodes keeps the troubleshooting scope small if something breaks.
Step 4: Open the official templates
In the ComfyUI template library, search for:
- MiniMax-H3-Image: Text to Image for generating images from text.
- MiniMax-H3-Image: Image Edit for uploading an image and describing the change in one sentence.
The image comes from ComfyUI's official template repository. A few things work differently from other models:
- It's fixed at 6 steps with no CFG, so there's no negative prompt. To keep something out of the picture, describe the absence in your positive prompt, like "an empty street with no pedestrians."
- Both templates start with a prompt expansion step that rewrites your short description into a detailed English one. In the text-to-image template it's off by default, so what you type is what gets used. In the image-edit template it's on by default.
- You can add reference images when editing. The first image is the one you're changing, and it's called
<Picture 1>. Add another Load Image node and connect it to the prompt generation and conditioning nodes, and that image becomes<Picture 2>. You can go up to 16. Use those names in your prompt to say which image provides what, for example: "put the jacket from <Picture 2> on the person in <Picture 1>."
Step 5: Start at 4MP, then scale up based on your VRAM
The text-to-image template opens with a 4MP square canvas, which is smaller than the model's native 16MP. That gives machines with less VRAM a chance to get a first run working. Once it works, use the resolution picker to increase the size.
If you run out of VRAM:
- Text to image: Set the resolution back to 4MP or lower.
- Image edit: Switch on the template's Resize Image/Mask node, set a smaller megapixel target, then run it again.
FAQ
Can MiniMax H3-Image run in ComfyUI? Yes. ComfyUI has supported it natively since 0.45.0. You don't need any custom nodes, and the template library includes official text-to-image and image-edit templates.
How much VRAM does MiniMax H3-Image need? The official template doesn't list a peak. It only says the 4MP default is there to help machines with less VRAM. Pruna's slimmed-down P-H3-Image gives a specific number: about 21 GB at peak when generating at 8MP.
How do I edit an image with MiniMax H3-Image in ComfyUI?
Open the Image Edit template, upload the image you want to change (that's your <Picture 1>), and describe the change in one sentence. To use other images as references, add more Load Image nodes and refer to them in your prompt as <Picture 2> and so on.
Why is there no negative prompt in the MiniMax H3-Image template? The model is fixed at 6 steps with a guidance strength of 1 and doesn't use CFG, so negative prompts aren't supported. Write what you want to avoid into the positive prompt instead.
Can MiniMax H3-Image make videos? No. H3-Image only generates still images. The image-to-video workflows you'll see in the community use MiniMax's H3 video model, which is a separate model.
Watch out
- Check your disk space first: The four files for the official version add up to about 55 GB.
- The license isn't a common open-source one: It's MiniMax's community license. Read it carefully before using the output commercially.
- Treat community low-VRAM workflows as your own risk: They often need third-party nodes and run slowly.
- If you only need a few images now and then, MiniMax's hosted service is simpler. See MiniMax H3-Image news.
Comments (0)
No comments yet. Start the conversation.