On 11 November 2026 MiniMax released the weights of its new image model, MiniMax H3 Image. It is a single model for both text-to-image generation and editing images from reference photos, and it draws natively at around 16 megapixels, with no upscaling step. ComfyUI has an official template for it, and on an RTX 3090 a full-resolution image takes 45 seconds.
What H3 Image is
H3 Image is built on MiniMax H3, the video model I have written about before. It shares the same architecture, only without the audio branch. At its core is a diffusion transformer with 33.1 billion parameters. Prompts are read by the same text encoder (Qwen3-VL) that video H3 uses, so if you already use video H3 in ComfyUI, you do not need to download the encoder again.
The model is distilled to 6 steps and needs neither a negative prompt nor CFG. It works in a range of roughly 4 to 20 megapixels and gives its best results around 16 MP. MiniMax recommends these sizes for common aspect ratios:
- 1:1 – 4096 × 4096
- 3:2 – 5024 × 3360
- 16:9 – 5472 × 3072
- 21:9 – 6272 × 2688
According to MiniMax, it follows prompts more precisely than any other open model. In the images this shows mostly in the details: skin, fur, fabrics and printed text look like a photograph, not a polished illustration.
How to run it in ComfyUI
Comfy-Org has published official templates for text-to-image (image_minimax_h3_image_t2i) and for editing (image_minimax_h3_image_edit). All you need is to update ComfyUI through the Manager; in testing the templates ran on version 0.45.3.
There are three files to download:
- the model itself in the int8 version used by the template – 21 GB,
- the VAE – 1.2 GB,
- the prompt enhancer (Qwen 3.5 9B) – 7.3 GB.
The text encoder (27 GB in int8) is shared with video H3. If you already have video H3, you only download the roughly 30 GB above. Everyone else will find it in the same Comfy-Org repository. There is also a smaller w4a8 version of the model, 12.6 GB, suited to cards with less memory.
On an RTX 3090 with 24 GB of memory, a 5024 × 3360 (16 MP) image takes 45 seconds once the model is loaded. That is how the cover image of this article was made. The first load of the model adds just under a minute.
You can write prompts in your own language
The model itself wants a detailed English description of the whole scene: who is in it, where they stand, the light, the materials and the lens. That is why the template includes a prompt enhancer. You write a short sentence in any language and the enhancer turns it into the long English description the model draws best from. The MiniMax API offers the same service under the name H3-Image-Context-IR.
When you want full control over the image, turn the enhancer off and write the description yourself. The template has a single switch for it. The model then draws only from your text.
Editing with up to 16 photos
The same model also edits images. It takes up to 16 reference photos, and in the prompt you refer to them as <Picture 1>, <Picture 2> and so on. MiniMax demonstrates this on several practical tasks:
- an empty room and nine photos of furniture combined into one furnished living room,
- a grey untextured 3D scene turned into a photorealistic film still,
- a drawn game map turned into the player's view from the spot the arrow points to.
For client work the key point is that a product, a character or a location only needs to be photographed, and the model places it into a new scene without any extra training.
A storyboard in a single image
The model can also draw a picture storyboard. In one prompt you describe the place, the characters and the action panel by panel, and you get a whole 3 × 3 sheet at once. The characters keep the same face, hair and clothes in every panel, and the space stays consistent across the shots. The model keeps the shot sizes (wide, medium, close-up) exactly as you write them in the prompt.
Thanks to the high resolution, each panel still has about one megapixel even in a large sheet. Such a sheet can go straight into the H3 video model as a guide, and it will shoot a continuous scene from it.
Text in the image
H3 Image writes legible text on posters, signs and packaging. MiniMax's examples even include whole boards with a logic puzzle in English and Chinese, where the model solves the puzzle correctly on the board. For poster, packaging or YouTube thumbnail designs this saves adding the text later in a graphics program.
Licence
The model does not use a standard open-source licence but its own MiniMax-H3-Image Community License Agreement. The public text excludes use in the European Union, the United Kingdom, South Korea and the USA. For these territories MiniMax offers a licence on request through its platform. Above an annual revenue of 1 million USD, and for running the model as a service for others, you need written permission from MiniMax. Before any commercial deployment, read the full licence; it is included with the model on Hugging Face.
Where to start
The fastest route is the official template in ComfyUI. Start at 4096 × 4096 with the enhancer on and a short prompt in your own language. Once you see what description the enhancer made of your sentence, you will know how to write your own detailed prompts. If you want to bring H3 Image into image or video production at your company, get in touch.
