MiniMax Image Generator

MiniMax H3 Image is an AI image model from MiniMax, the lab behind the MiniMax-H3 video model. MiniMax released it on November 10, 2026, with open weights on Hugging Face as MiniMaxAI/MiniMax-H3-Image. A single model both creates images from a written prompt and edits images you give it, with up to 16 reference images in one run.

What sets it apart is size. MiniMax H3 Image renders about 16 megapixels natively, with no upscaler: 4096 x 4096 for a square image, 5472 x 3072 at 16:9 and 6272 x 2688 at 21:9. It also follows long, detailed instructions closely, and MiniMax calls its instruction following the strongest among open-source models. Strong photorealism, rich fine detail and exact text in quotes round out what it does well.

Mage runs MiniMax H3 Image in your browser, next to the rest of its image, video and audio models, so there is nothing to download and no graphics card to own. Mage is an independent creative platform, not part of MiniMax. What it adds is everything around the image: unlimited plans, saved Characters and References you call by @handle, LoRAs on top of the model, and apps for fixing and relighting what you made.

An upscaler enlarges a finished picture by guessing at the detail it lacks. MiniMax H3 Image does not need one: it renders about 16 megapixels natively at 4K. A square image is 4096 x 4096, a 16:9 frame is 5472 x 3072, and a 21:9 ultrawide is 6272 x 2688, all drawn by the model itself. Fine detail such as skin, fabric, leaves, stone and small print comes from the model itself, not from a second pass.

On Mage it runs at two sizes. 2K is about 4 megapixels and takes about 3.5 seconds, which makes it the size for trying ideas, testing a prompt and running variations. 4K is the full 16 megapixels and takes about 13 seconds: switch to it for prints, posters, wallpapers, product photos and anything people will zoom into. Both sizes come in nine aspect ratios, from a 21:9 ultrawide to a tall 9:21 frame, with square, portrait, landscape and phone shapes in between.

A simple workflow covers most jobs: draft at 2K until the prompt is right, then run the final at 4K. The MiniMax H3 Image model page has samples, every setting and example prompts.

• Drafts, prompt tests and variations: 2K
• Prints, posters and wallpapers: 4K
• Ultrawide monitors and banners: 21:9 at 6272 x 2688
• Phone wallpapers and stories: 9:16 at 4K

MiniMax H3 Image is built to follow long, detailed instructions, so give it a full description rather than a list of keywords. Describe the whole image: the subject and what it is doing, the setting, the composition, the light, and the style or medium. Then be specific where it matters. Say how many objects there are, where each one sits, and what color and material it is. The more complete the brief, the less the model has to guess.

The model is distilled to six steps, so there is no negative prompt, step count or guidance scale to tune. Everything you want, and everything you want left out, goes into the prompt itself: write "no other text" or "plain grey background" into the sentence. A seed setting is there when you want to compare two versions of a prompt from the same starting point.

For text in the image, put the exact words in double quotes and say where they go and how they look: the headline "ISLAND HOPPING" in bold cream capitals across the top. Keep each line short, and when the words sit on something in the scene, say what it is, such as a chalkboard menu, a shop sign or a book cover.

The MiniMax H3 Image prompting guide goes further, with worked examples for posters, edits, and scenes built from many images.

• Describe the whole image in full sentences
• Count and place the important objects
• Quote any text that should appear
• Write exclusions into the prompt

MiniMax H3 Image edits with the same model it generates with. Upload a photo and write the change as an instruction: say what to keep, what changes, and how the finished picture should look. For example, "keep her face, hair and pose from image 1, change the denim jacket to a yellow raincoat, and put her on a rainy street at night". Name the thing you mean instead of writing "it".

One run takes up to 16 reference images, counting uploads and the saved Characters and References you mention by @handle. The prompt refers to them in order as image 1, image 2 and so on, so you can say exactly what to take from each: the room from image 1, the sofa from image 2, the lamp from image 3 and the rug from image 4. That suits jobs that bring many pieces together, such as a furnished room, a product line-up or a group of characters in one scene. When the same face or product comes back again and again, save it once and call it by its @handle instead of uploading it every time.

For a pixel-level fix in one spot, such as a stray hand or a wrong letter, brush over it in Inpaint. To change the light of a finished image, open Relight, and for more ways to edit a picture with words, see the AI image editor page.

Because MiniMax H3 Image has open weights, its community can train LoRAs for it: small add-on files that push the model toward a style, a character or a look. On Mage, MiniMax H3 Image LoRAs work on top of the model. Paste the CivitAI or Hugging Face link of a LoRA on the Import page, and it appears in the LoRA picker for MiniMax H3 Image.

Importing is part of the Pro Plus and Max plans. The LoRA image generator page explains how to set a weight, use trigger words, and stack LoRAs without them pulling against each other.

Open weights mean anyone can download MiniMax H3 Image from Hugging Face and run it themselves. In practice, a model that renders 16 megapixels at a time takes a powerful graphics card, the model files and a working setup, and the setup grows once reference images and LoRAs come into it.

On Mage it runs in the browser on any device, a phone included, with reference images, saved Characters and LoRA support already in place. You get what makes an open model worth using, including LoRAs from its community, and skip the install. And because Mage carries image, video and audio models in the same account, a still you make here can become the first frame of a video without leaving the site.

Mage publishes side-by-side comparisons of MiniMax H3 Image with four popular image models: MiniMax H3 Image vs Nano Banana 2.1, MiniMax H3 Image vs GPT Image 2.5, MiniMax H3 Image vs Seedream 5.5 and MiniMax H3 Image vs Flux 2. All five models are in the same model menu on Mage.

On a paid plan, a MiniMax H3 Image run needs no Gems when its size fits the plan. Pro, at $30 a month, makes MiniMax H3 Image unlimited at 2K. Pro Plus, at $60 a month, adds unlimited 4K, the full 16 megapixels. Unlimited runs use the standard queue.

Without a plan, or at a size your plan does not cover, MiniMax H3 Image runs on Gems. A 4K image costs more than a 2K one, and each reference image adds a few Gems. Gem packs start at $10 for 10,000 Gems, and every paid plan includes monthly Gems, 1,500 on Pro.

Compare what each plan covers on the membership page, or see every model that runs without limits on the unlimited AI image generator page.