正文 · AI 翻译
“CAD Disassembly”
By Jay, reporting from Aofei Temple, QbitAI | WeChat account QbitAI
By Jay, reporting from Aofei Temple, QbitAI | WeChat account QbitAI
Well, MiniMax really has figured out multimodal AI 😭.
After its open-source flagship video model H3 won widespread praise, its latest image model, H3-Image, has once again taken the internet by storm.
The hottest trend is “CAD disassembly”: give H3-Image a reference image, and it can lay out all the parts of a little figurine for you...
Someone even made a CAD drawing of a Unitree robot dog with it.
In short, it can take just about anything apart and label the parts for you.
I tried it myself, and I have to say, it really can do something...
Besides that, overseas users have found it can handle other kinds of complex image-generation tasks too, such as games.
If you hadn't told me, I would have thought this was made by programming with GPT 😭.
If you just need simple assets for e-commerce, that's even easier.
But what I didn't expect was that this thing apparently has both artistic and technical chops??
Alongside its intelligence, H3-Image has a pretty good aesthetic sense too.
Oh, and one more thing.
You're telling me an image-generation model of this quality is actually open source???
What more is there to say?
I hereby declare MiniMax the “open-source god” of multimodal AI (just kidding).
Goodbye to a Plastic Aesthetic Sense
In the past, the first thing that came to mind when we mentioned AI-generated images was shiny faces, skin as smooth as peeled eggs, vacant stares, and an overwhelming uncanny-human feel.
In short, they had a strong “AI look.”
GPT Image 2 has improved enormously on this front, but after using it a lot, the uncanny-valley feeling still creeps in.
With H3-Image this time, I feel that problem has basically been solved.
I was stunned when it finished generating. Seriously, who could tell this was AI??
Most importantly, it renders faces down to the pores, freckles, and even fine facial hair.
It also handles details like moisture and dust on the mirror and the halo made by a flash hitting the lens remarkably well. It really looks like something you'd scroll past on the Xiaohongshu homepage.
You can see plenty of facial detail even from below. Take the fashion-magazine cover below, where a table becomes an Instagram-worthy photo: look closely at the person's gaze and the subtle pull of their facial muscles. Both feel entirely plausible in a real social interaction.
That was also my strongest impression while testing H3-Image: it gives fresh interpretations to many prompts I'd previously tried with older models.
It really does have fairly broad knowledge of the world.
Subjects other than faces are even easier for it.
For example, turn Rio de Janeiro's landmarks into a travel journal in the style of a three-dimensional paper sculpture.
The miniature scene and the depth of the light and shadows—hmm, somehow it feels like a Fable 5 trailer.
And sure enough, it showed its little tell again!
See that hand in the lower left? It added a few moles near the thumb joint!!
Zoom in and the illusion still holds up: the texture of the skin on the fingers, the joint creases, and even the natural sheen and texture along the edges of the nails are all rendered clearly.
Interesting.
I think H3-Image's lack of that AI look may largely come down to these deliberate “imperfections.”
Any work created by a human is bound to be full of randomness. An excessive pursuit of perfection actually runs counter to how the human brain is used to taking in information.
But that also makes me a little uneasy.
I had stubbornly believed that aesthetic judgment was something machines couldn't learn.
Now it seems that, if you can figure out the tricks behind it, many qualities humans pride themselves on may not be so mysterious after all...
Let me show you another demo: I used H3-Image to make a France travel poster.
How should I put this? I'm almost embarrassed to judge it; my artistic knowledge isn't enough to make sense of it anymore.
Sigh.
SOTA-Level Intelligence
Let's try some tasks I can actually understand.
When it comes to multimodal models, I've always had 2 criteria: aesthetic sense and intelligence.
Aesthetic sense is what I tested above. It's rather hard to pin down, so I can only talk about my subjective impressions.
But intelligence is something we can dig into.
For instance, H3-Image has a rather unusual skill: understanding the physical world.
You can use it to generate CAD drawings.
Yes: give it an image of what a product looks like, and it can automatically work out the standard 3 orthographic views—
Front, side, and top views, all present.
Friends, I'm weak at the knees 😭😭.
And that's not all.
Besides CAD drawings, you can also ask the model to “explode” an object.
Ask H3-Image for an exploded mechanical diagram, and it can follow the assembly logic accurately, separate the parts into clear layers, and even label the component callouts.
That undoubtedly requires a deep understanding of three-dimensional geometry and spatial relationships.
Of course, for images with such clear functional requirements, following instructions is also very important. Simply “drawing” them isn't particularly hard; the main test is the model's ability to plan and reason.
Otherwise, it couldn't make sense of what users mean in all those prompts.
Of course, for serious production work, the bar has to be set very high: commercial uses must convey information accurately.
So beyond spatial understanding, text layout is also extremely important.
So I tried another thing—
I generated a final exam paper for advanced calculus.
It was ready in just over 10 seconds.
Limit notation, complex integrals... In this kind of dense academic layout, somehow there wasn't a single garbled character.
That must have been too easy, so I came up with a brilliant prompt:
Complete this exam paper.
Complete this exam paper.
When I opened it, I could practically see my chair explode—
What do you mean an image-generation model is better at calculus than I am???
Drawing on the advanced calculus knowledge I barely managed to graduate with, I looked it over and found that the answers were actually right—and it even showed the steps!??
For a moment, I was speechless.
Of course, we can't leave out the usual task: generating assets.
Isn't it fashionable now to say coding is swallowing everything? Astra is taking on 3D, and Opus 5.5 is taking on video.
But I've always felt that much of this is marketing talk. Why would I use a text model for tasks so far removed from text?
I can only say that the text-only field does have plenty of resources and has developed quickly.
Starting from first principles, I still believe in specialized models trained for their respective fields. Their ceiling will be higher, and their costs more realistic.
And a specialized model like H3-Image is naturally suited to its specialty. There's no contradiction at all: it's an excellent fit for a harness that lets an agent call it directly as a tool.
This seems to be an emerging trend: a single model doesn't necessarily have to do everything. Connecting specialized models through a general-purpose model may produce greater overall value.
If all you're doing is generating an asset for display, such as the dinosaur toy below, a specialized image model is clearly the better choice. There's no need to spend ages on computer use.
After all, basic physical concepts such as lighting, shadows, and geometry are no longer a problem for today's image models.
Snacks are no problem either. Watch me take Chinese latiao spicy strips global.
It can even generate 3D game scenes. Everything's turned upside down 😭.
Give it a long tactical instruction—cover, suppressing fire, flanking maneuvers—and it can immediately build a 3D combat scene with convincing depth and plausible occlusion.
It follows game terminology closely and could almost be used directly in concept design.
Why should I bother with Astra Ultra, then? If all I'm doing is looking at a demo, why not just generate a few images?
(Just kidding.)
The Multimodal “Open-Source God”
That's basically it. Beyond all that, as a flagship image model, H3-Image has a few more perks.
Most importantly, H3-Image pushes resolution to 16MP (4096×4096 pixels), or native 4K.
I'm tearing up, friends. At last I don't have to generate an image with GPT and then spend ages improving it with Nano Banana 🥹.
And 4K generation is incredibly fast. For images this large, it felt like just 10–20 seconds on the web interface.
That said, who wouldn't want to enjoy a game of Snake while waiting for GPT to dawdle? (Just kidding.)
I digress.
Besides these, H3 has another rather unusual specification—
It supports input of up to 16 reference images.
You can fit a huge amount in. It's a perfectionist's dream, giving creators unprecedented freedom with character consistency and continuity across complex storyboards.
Everyday 4K creation is now truly within reach.
The result is that you can casually toss in a few reference images, take a sip of water...
And get an interior design rendering of this quality.
Who even needs Blender now 😭?
But what I never expected was this.
H3-Image is actually open source???
After previously open-sourcing the SOTA-level video model H3, MiniMax has done it again, giving developers worldwide another freely customizable SOTA foundation.
The multimodal open-source god. Enough said.
Of course, even if you don't feel like dealing with deployment, API calls are dirt cheap.
They cost just 50% to 70% as much as those of mainstream competitors, and 16MP requests are another 50% off during the first week (get on it, folks!!).
By this point in my testing, I suddenly had a feeling—
Multimodal generation may be approaching its own coding moment.
In the past, AI image generation was often “reduced to a toy” in commercial settings, mainly because the details didn't stand up to scrutiny and the plastic look was too pronounced. What client would be happy with that?
But now H3-Image has achieved 3 things—
Less of an AI look, an aesthetic sense that generalizes, and logic.
When those conditions are met, the market for multimodal applications will certainly be no smaller than the one for coding.
From high-precision industrial packaging design and commercial illustration to multilingual materials for cross-border marketing, image models can greatly shorten traditional workflows.
Add the advantages just mentioned—speed, open source, and low prices...
It's hard to imagine what kind of future that will bring.
But finally, let's turn our attention back to today's AI industry.
While the attention of almost every AI company in the world is focused on text alone, MiniMax continues to devote a large share of its efforts to the multimodal world.
That's long been a defining trait of the company. It bet on multimodal AI from the start, and now its efforts are finally bearing fruit.
People say multimodal AI doesn't matter for AGI, and that coding is where the business model is...
But I've always thought that if the definition of AI has room only for text, that future would be far too dull.
From H3, which reshaped video generation, to H3-Image, which now makes comprehensive strides in logic, structure, and aesthetics.
The wildfire of Chinese open-source AI has now truly begun to spread into multimodal AI—
And keep spreading.