MiniMax H3 Image Guide: 4K Images, 16 Reference Images and Open Weights for Marketing Teams
MiniMax H3 Image makes 4K images and edits with up to 16 references, and its weights are open. What MiniMax and fal document, how it is billed, and where it fits a marketing workflow.

Most image models make you choose: a sharp 4K render, or a quick draft, or a model that can combine lots of reference photos. MiniMax H3 Image, released on November 11, 2026, goes after all three: 4K output, edits that take up to 16 reference images, and a Highspeed mode for drafts. Its weights are also open. This guide covers what MiniMax and fal document, how it is billed, and where it fits an ad or content workflow.
One thing to know up front: MiniMax H3 Image is not in MITPO Creative Studio yet. fal bills it by the token and publishes no per-image price, so we want to measure the real cost of a picture before we put a credit price on it. Everything below is what the vendor and the host document, and we have not judged the output quality ourselves.
Key takeaways
- What it is: MiniMax's image model, built on its H3 video model, for text-to-image, reference-based generation and editing.
- Size: up to 4K (4096 by 4096, about 16 megapixels) according to MiniMax. fal accepts any output between 655,360 and 20,000,000 pixels.
- References: up to 16 input images per edit, cited in the prompt as Image 1, Image 2 and so on.
- Speed: MiniMax says a 4K image takes under 10 seconds, and describes a Highspeed mode at 3 to 4 times the speed of mainstream alternatives.
- Billing on fal: by token, at $23 per 10 million input text tokens, $30 per 10 million input image tokens and $76 per 10 million output image tokens.
- In MITPO: not available yet.
What MiniMax says it does
From MiniMax's launch post (November 11, 2026), as the vendor's own claims:
- An open-weight image model. The weights are published on Hugging Face as MiniMax-H3-Image, as with the H3 video model.
- Text-to-image, reference-based generation and editing in one model.
- Up to 16 reference images.
- Up to 4K output, 4096 by 4096.
- Speed: a 4K image in under 10 seconds, standard mode 1.5 to 2 times as fast as mainstream alternatives and Highspeed mode 3 to 4 times as fast.
- Price: MiniMax says 50 to 70 percent of comparable mainstream offerings, with 50 percent off for 4K and above during launch week.
The post gives no independent benchmarks, and the exact prices are shown only in a chart. Check the license text on Hugging Face before using the weights commercially.
What fal documents
| Setting | What the schema says |
|---|---|
| Prompt | Up to 32,000 characters |
| Prompt expansion | On by default: a language model rewrites your prompt into a long, detailed description first. If the rewrite fails, the request fails |
| Output size | Any exact size from 655,360 to 20,000,000 pixels, aspect ratio between 1:3 and 3:1 |
| Images per request | 1 to 10, all generated from one expanded prompt |
| Output format | PNG (default), JPEG or WebP |
| Edit endpoint | Up to 16 input images, each side 128 to 8192 pixels, cited as Image 1, Image 2 and so on |
| Edit size | "auto" keeps the first image's aspect ratio, within a 2048 by 2048 area |
There is a separate Highspeed endpoint with the same inputs.
How it is billed on fal
fal bills actual token usage rather than a flat price per image:
| Tokens | Rate |
|---|---|
| Input text | $23 per 10 million |
| Input images (edits) | $30 per 10 million |
| Output image | $76 per 10 million |
Prompt expansion is counted as input text. fal does not publish a per-image price, and the cost of one picture depends on its size, because a bigger image uses more output tokens. For a campaign budget, run a few real requests at the sizes you plan to use and read the cost from your fal usage page before you scale up.
Where it fits a marketing workflow
- Hero images at large sizes. A 4K canvas can be cropped for a banner, a poster and a social post from one render.
- Product shots built from references. Feed in the product, a background and a style example, and cite each by number in the prompt.
- Campaign variants. Ask for several images from one prompt, then pick.
- Fast drafts. Use Highspeed to explore ideas, then re-render the winner in the standard mode.
- Self-hosting. Open weights mean a team with the hardware can run it privately. MiniMax has not published local memory requirements in the post we read.
How to prompt it
Because prompt expansion is on by default, a short prompt gets filled out for you. Turn it off when you want the model to follow your wording exactly.
1. Clean product hero
[Product] on a [surface], [setting], soft diffused light, shallow depth of field,
generous empty space on the left for a headline, photographic, no text in the image.2. Multi-reference composition
Use Image 1 as the product and keep its shape and colour exactly. Use Image 2 for the
lighting and mood. Place the product on the surface from Image 3. Keep everything
else simple and uncluttered.3. Targeted edit
Put [new element] on [subject] in Image 1 and keep everything else the same:
composition, lighting and background.Limits to plan around
- Vendor claims. The speed, resolution and price claims are MiniMax's own and unverified by us.
- Token billing is hard to forecast. There is no per-image price to plan on, so test first.
- Prompt expansion can change your wording. Turn it off for exact copy or layout instructions.
- License. Open weights come with a license. Read it for your use and your region before you rely on the weights.
- Not in MITPO yet. We have not switched it on in Creative Studio.
Trying it with MITPO
MiniMax H3 Image is not in Creative Studio today, so to compare models for a brief, use the ones that are, such as Seedream 5.5 or GPT Image 2.6, and see the pricing page for plan limits. Once we have measured what a picture costs, we will add it with a credit price.
FAQ
What is MiniMax H3 Image?
It is MiniMax's open-weight image model, built on its H3 video model, for text-to-image, reference-based generation and editing. MiniMax released it on November 11, 2026.
How many reference images can it use?
Up to 16, on both MiniMax's launch post and fal's edit endpoint.
What resolution does MiniMax H3 Image support?
MiniMax says up to 4K (4096 by 4096). fal accepts any output between 655,360 and 20,000,000 pixels with an aspect ratio between 1:3 and 3:1.
How much does it cost?
fal bills by token: $23 per 10 million input text tokens, $30 per 10 million input image tokens and $76 per 10 million output image tokens. It does not publish a per-image price. MiniMax says its own prices are 50 to 70 percent of comparable offerings.
Is MiniMax H3 Image open source?
MiniMax has released the weights on Hugging Face. Check the license terms for your use before relying on them.
Can I use it in MITPO?
Not yet. We will add it once we have measured its real cost per image.
Put this to work in MITPO
Open the demo with no account, or create a free account and start with your own brand.
MITPO Editorial & Research Team
Operator-verified playbooks benchmarked on active 2026 marketing workflows.
Further Reading
Related Articles

Design Trends That Will Dominate 2026
From neo-brutalism to glassmorphism 2.0, explore the visual styles that are defining the aesthetic of the next web generation.

Generative UI: The Future Where Interfaces Design Themselves
Imagine an app that changes its entire layout based on who is using it. Generative UI is making static interfaces a thing of the past.

Top 5 Mistakes Startups Make With Their Landing Pages
Five common landing page mistakes, each with a fix and the evidence behind it, including two real A/B test case studies.