MiniMax Open-Sources Image Model H3-Image, Supporting Native 4K Generation
MiniMax has open-sourced its image model H3-Image, which supports native 4K generation at 16MP (4096×4096), with image output on the web taking roughly 10-20 seconds in the author's experience. In hands-on testing, the author found it approaches real photographs in details such as facial pores and fine hairs, and specular water highlights; it can also derive orthographic three-view drawings and mechanical exploded views from product appearance images, and correctly render complex integral expressions. API pricing is 50% to 70% of mainstream competitors, with an additional 50% off 16MP requests in the first week.
Why read this
SelectedAI score 83
The author presents first-hand hands-on testing of H3-Image's performance in detail texture, technical drawing reasoning, and native 4K, along with open-source and API pricing information.
Topics
Body · AI translation
Multimodal "Source God"! MiniMax open-sources SOTA image generation model, native 4K, two dimes per image
After the open-source video flagship H3 was widely acclaimed, the latest image model—H3-Image—has once again swept the entire internet.
The hottest thing is "CAD teardown": toss it a reference image, and H3-Image can directly unbox your little figurine for you...
Beyond that, overseas netizens who tested it found that other types of complex image generation tasks can also be handled, such as—games.
This this this, if you hadn't said anything, I'd have thought this was made by GPT coding 😭.
If it's just making some simple assets for e-commerce, it's even easier.
But what I didn't expect was that this guy is actually good at both arts and sciences??
On the premise of guaranteeing intelligence, H3-Image's aesthetics also have something going on.
Oh right, one more thing.
An image generation model of this quality, and you're telling me it's actually open source???
Then what else is there to say.
I hereby declare that MiniMax is the "Source God" (not) of the multimodal field.
Saying goodbye to plastic "aesthetic sense"
GPT Image 2 made huge progress in this regard, but after using it a lot, the uncanny valley effect still surges up.
This time with H3-Image, I feel it has basically solved this problem.
The most important thing is faces—pores, freckles, even fine hairs are all drawn out for you.
In addition, details like the water sheen and dust on a mirror, and the halo formed by a flash hitting the lens, are handled incredibly well—it really looks like something you'd scroll past on the Xiaohongshu homepage.
Even from a low angle you can see plenty of facial details. For example, in the fashion magazine cover below, which looks like an Instagram photo of a table, if you look closely at the person's gaze and the subtle facial muscle tension, they all fit real-world social logic very well.
This is also the biggest impression I had when testing H3-Image: many prompts I had thrown at older models, H3-Image can give a completely new interpretation of.
Indeed, its world knowledge is fairly rich.
As for subjects other than faces, it's even more of a sure thing.
For example, turning Rio de Janeiro's city landmarks into a travel journal in a 3D paper-cut style.
Miniature landscape, depth of light and shadow, hmm, it somehow has a whiff of a Fable 5 promo trailer.
And sure enough, it showed its telltale flaw again!
Look at this hand in the lower left corner, see it? It added a few moles near the thumb joint!!
Zoom in and it won't give itself away at all—the skin texture on the fingers, the joint wrinkles, even the natural sheen and texture at the edges of the nails, are all reproduced clearly.
Interesting.
I think the reason H3-Image has so little AI flavor may largely depend on these deliberately created "flaws."
Any work subjectively created by humans is inevitably full of randomness; excessively pursuing perfection actually goes against how the human brain is used to receiving information.
But this also makes me a little uneasy.
I have always stubbornly believed that aesthetics is not something machines can learn.
Now it seems that as long as you can figure out the tricks involved, many qualities humans are proud of may really not be that mysterious...
Let me show you another demo—I used H3-Image to whip up a French travel poster.
How should I put it? I really don't feel qualified to comment; my humble artistic literacy can no longer make out what's going on.
As far as multimodal models are concerned, I have always had two criteria—aesthetic sense and intelligence.
Aesthetic sense is what I tested above; it's rather mystical, so I can only talk about subjective feelings.
But intelligence is something we can discuss.
For example, H3-Image has a very special skill point—world understanding ability.
You can use it to generate CAD drawings.
Yes, you give it an exterior image of a product, and it can automatically derive standard orthographic three-view drawings—
Have H3-Image draw an exploded mechanical teardown diagram, and it can accurately follow the mechanical assembly logic, blow the parts apart layer by layer clearly, and even label the component leader lines for you.
Behind this is undoubtedly a deep understanding of 3D geometry and spatial relationships.
Of course, for images with obvious functional attributes like these, instruction-following ability is also very important. Just "drawing" isn't that difficult; the main test is the model's planning and reasoning ability.
Otherwise, it simply wouldn't be able to figure out what the user's pile of prompts is saying.
Of course, if you really want to hit serious productivity scenarios, the standards must be set very high; commercial use must accurately convey information.
Limit symbols, complex integral expressions... this kind of high-density academic typesetting, and not a single character came out garbled.
With my barely-passing advanced math knowledge, I took a rough look and found that the answer is actually correct, and there are even solution steps!??
For a moment I was speechless.
Of course, in its old line of work, asset generation can't be left out.
Isn't it trendy now for Coding to eat everything? Astra reduces dimensions to 3D, Opus 5.5 reduces dimensions to video.
But I've always felt that a lot of this is marketing talk. For these tasks that have nothing to do with text, why would I use a text model?
I can only say that the pure-text field indeed has more resources and develops relatively fast.
Starting from first principles, I still rather believe in the vertical-model path: train each field separately, and the ceiling will be higher while the cost is more realistic.
Moreover, vertical models like H3-Image are born specialists—there's no contradiction at all, and they're very suitable for plugging into a Harness so Agents can directly Tool Call.
This also seems to be a trend that has started emerging recently—it doesn't necessarily have to be a single model that reduces everything to one dimension; using a general model to connect various vertical models yields higher overall utility.
If it's just about generating assets for display, like the dinosaur toy below, a vertical image model is clearly the better solution—there's no need to spend ages with Computer Use.
After all, with today's image models, basic physical common sense like lighting and geometry is no longer a problem.
Snacks and the like are no problem either—just watch how I take Chinese latiao global.
It can even generate 3D game visuals—the world's turned upside down, I swear 😭.
Give it a long tactical instruction—cover, suppressive fire, flanking maneuvers—and it can immediately build you a 3D combat scene with a real sense of depth and reasonable occlusion relationships.
It follows game terminology strictly and can almost be used directly in concept design.
Why would I even bother with Astra Ultra then? It's all just looking at demos anyway—just generate a few images and be done with it.
(not serious)
The multimodal "Source God"
That's basically it. Beyond that, as a flagship image model, H3-Image has also released some small perks.
Most importantly, H3-Image pushed the resolution all the way to 16MP (4096×4096 pixels), i.e., native 4K.
I'm tearing up, folks—I finally don't have to generate an image with GPT and then toss it to Nano Banana for ages of optimization 🥹.
And 4K generation is super fast too—for an image this big, my subjective experience on the web version is just 10-20 seconds.
Still, who wouldn't want to enjoy a nice game of Snake while waiting for GPT to dawdle (not serious).
It can hold a huge amount—a godsend for perfectionists—giving creators unprecedented freedom in character consistency and continuity across complex storyboards.
Following the previously open-sourced SOTA-level video model H3, MiniMax has acted again, giving global developers another freely customizable SOTA foundation.
Multimodal Source God—need I say more.
Of course, even if you don't feel like messing with deployment, API calls are dirt cheap.
Only 50% to 70% of mainstream competitors' prices, and 16MP requests get another 50% off in the first week (go hard, brothers!!).
Having tested this far, I suddenly have a premonition—
Multimodal generation may be ushering in its own Coding moment.
But now, H3-Image has accomplished three things—
When these conditions are met, the multimodal application market will absolutely be no less impressive than programming.
From high-precision industrial packaging design and commercial illustration to multilingual cross-border marketing materials, the emergence of image models can greatly compress the cycles of traditional workflows.
Stack on the buffs just mentioned: speed + open source + low price...
What kind of future this will be is hard to imagine.
When the eyes of global AI companies are almost all focused on pure text, MiniMax still devotes a large part of its energy to the multimodal world.
This has also been a longstanding characteristic of the company—betting on multimodality from the very beginning, and now finally seeing emergence.
Everyone says multimodality isn't important for AGI, that only Coding has a business model...
From H3, which rewrote the video generation landscape, to H3-Image, which now makes comprehensive breakthroughs in logic, structure, and aesthetics.
This wildfire of domestic open source has thoroughly begun to spread into the multimodal field—
spreading.