Multimodal "Source God"! MiniMax open-source SOTA image generation model, native 4K, just 1.2 yuan

Jay from 凹非寺 Quantum Bit | Official Account QbitAI

Bad, really let MiniMax take mult模态 to the next level😭。

Following the positive reception of the open-source video flagship H3, the latest image model——H3-Image, has once again swept the internet。

The hottest one is 「CAD Disassembly」, just drop a reference image, and H3-Image can directly open the box for your figurine。。。

Also, there are CADs forYuShu Robotics Dog。

In short, everything can be disassembled, and it can also label the parts for you。

I myself tested it, you wouldn't believe, it actually has some stuff……

In addition, overseas users found that it can handle other complex image generation tasks, such as——game。

This this this, you didn't say, I thought it was made by GPT programming 😭。

If it's just for simple assets used in e-commerce, it's even easier。

But what I didn't expect was that this guy is actually both literary and scientific? ?

Under the guarantee of intelligence, H3-Image's aesthetics also have some stuff。

By the way, there's another thing。

This quality of image generation model, you're telling me it's open-source? ? ? That's it。

I declare, MiniMax is the 「source god」 in the mult模态 field

（bushi）

。 Farewell to plastic 「aesthetics」

In the past, when we mentioned AI-generated images, the first thing that came to mind was: oily skin, smooth as an eggshell, vacant eyes, extremely fake human look.

In short, it's a very strong 「AI taste」。

GPT Image 2 has made significant progress in this aspect, but after using it a lot, the uncanny valley effect still comes up。

This time's H3-Image, I feel, should have basically solved this problem。

After generating the images, I was stunned, not brother, who can tell this is AI? ?

The most important thing is the face, pores, freckles,

even the fine hairs

are all drawn out。 In addition, details like water droplets and dust on mirrors, and the halo formed by flashlights on the lens, are handled extremely well, just like what you see on the homepage of Xiaohongshu。

Even from a仰视 angle, you can see a lot of facial details, for example the following image

table which is an Instagram photo

of a fashion magazine cover, if you look closely at the person's eyes and the subtle facial muscle movements, they are all very in line with real social logic。

This is the biggest feeling I had when testing H3-Image, many of the prompts I used to give to old models, H3-Image can give a brand new interpretation。

Indeed, the world knowledge is quite rich。

As for subjects other than faces, it's even more skilled。

For example, creating a travel journal in a 3D paper sculpture style of Rio de Janeiro's city landmarks.

Micro landscapes, light and shadow depth, hmm, it feels like the flavor of a Fable 5 promotional video.

And indeed, it's revealed the chicken feet again!

Look at this hand in the lower left corner, see it? It added a few moles near the thumb joint!

When you enlarge it, it won't show any flaws, the skin texture on the fingers, the folds around the joints, even the natural luster and texture at the edges of the nails are all clearly reproduced.

Interesting.

I think H3-Image has so little AI flavor, it's probably largely due to these intentionally added 'flaws'.

Any work created by human subjectivity must be filled with randomness, and pursuing perfection too much is actually not in line with the way the human brain receives information.

But this doesn't stop me from feeling a bit scared.

I've always stubbornly believed that aesthetics isn't something a machine can learn.

Now it seems that as long as you can figure out the tricks behind it, many of the qualities that humans take pride in might not be as mysterious as they seem...

Let me show you another demo. I made a French travel poster with H3-Image.

Honestly, I don't know how to evaluate it. My artistic sensibility is completely unable to grasp the essence.

Ugh.

SOTA-level intelligence

Let's do something that I can understand.

In terms of multimodal models, I've always had two evaluation criteria - aesthetic sense and intelligence.

Aesthetic sense is what I tested above, it's quite abstract, and can only be discussed in terms of subjective feelings.

But intelligence can be talked about more concretely.

For example, H3-Image has a very special skill - world understanding ability.

You can use it to generate CAD drawings.

Yes, it's the ability to automatically generate standard orthographic three-view diagrams from a product's exterior image -

front view, side view, top view, none missing.

Everyone, I'm really exhausted 😭😭.

It's not over yet.

Besides CAD, you can also make this model 'explode'.

Let H3-Image draw an exploded view of a mechanical structure, and it can accurately follow the assembly logic of the machine, exploding the parts in layers and clearly showing them to you, even labeling the component wires for you.

This clearly requires a deep understanding of three-dimensional geometry and spatial relationships.

Of course, for images with such obvious functional features, instruction following ability is also very important, just 'drawing' isn't actually that difficult , the main challenge is testing the model's planning and reasoning capabilities.

Otherwise, you'd never understand what the user's bunch of prompt words are talking about.

Of course, if we are really talking about serious production scenarios, the standards will definitely be set very high, and commercial use must accurately convey information. So, besides spatial understanding, the dimension of text layout is also very important. So, I did this thing again — I generated a high mathematics final exam paper. It was generated in just a few seconds. Limit symbols, complex integral formulas... this high-density academic layout, surprisingly, didn't have any character corruption. Probably because it was too simple, I came up with a genius prompt: Complete this exam paper. When I opened it, I felt like my chair exploded — What does it mean that an image generation model has a higher level of calculus than me?? With my barely passed calculus knowledge, I took a quick look and found that the answers were correct, and there were even step-by-step solutions!? At this moment, I was speechless. Of course, my specialty, asset generation can't be missing either. Now isn't it popular for coding to eat everything? Astra reduces 3D, Opus 5.5 reduces video. But I always think that a lot of these are just marketing slogans. These tasks have nothing to do with text, so why would I use a text model? It can only be said that the pure text field has more resources and is developing faster. Starting from the first principles, I still believe that the vertical model approach is better. Training separate models for each field, the ceiling will be higher, and the cost will be more realistic. Moreover, vertical models like H3-Image are naturally specialized, and there's no contradiction at all, making them very suitable for integration with Harness, Letting the Agent directly Tool Call. This is also a trend that has recently started to appear. It doesn't necessarily mean that a single model should reduce everything. By using a general model to connect various vertical models, the overall utility will be higher. If it's just about generating assets for display, like the dinosaur toy below, the vertical image model is clearly a better solution. There's no need to spend half a day on Computer Use. After all, the current image models have already mastered basic physical knowledge such as lighting and geometry. Well, well, well. Small snacks are also fine, and let me see how I can Make Chinese spicy noodles go global. Well, well, well. Even 3D game scenes can be generated, which is really reversing the situation. 😭 With a long text tactical instruction — what kind of cover, fire suppression, flanking, it can immediately build a 3D combat scene with real depth and reasonable occlusion relationships. Strictly following game terminology, it's almost ready to be used in concept design. Well, well, well. I don't even need to use Astra Ultra anymore. After all, it's just a demo, generating a few images is enough. (bushi) Multimodal "Source God" Basically, that's it. In addition, as a flagship image model, H3-Image also released some small benefits.

Most importantly, H3-Image directly boosted the resolution to 16MP (4096×4096 pixels) , which is native 4K.

I went off track.

In addition to these, H3 has another special parameter —

supports up to 16 reference images as input.

It can take in a huge number of images, which is great news for perfectionists. It gives creators an unprecedented level of freedom in terms of character consistency and continuity of complex scenes.

4K daily creation, truly becoming mainstream.

The result is, you can just throw in a few reference images, take a sip of water...

and get this quality of interior design renderings.

Who needs Blender anymore? 😭

But what I never expected was.

H3-Image, to be open-source ? ??

Following the open-sourcing of the SOTA-level video model H3, MiniMax once again made a move, offering the global developers another customizable SOTA foundation.

Multimodal God, no need to explain.

Of course, even if you don't want to bother with deployment, API calls are super cheap.

Only 50% to 70% of mainstream competitors' prices, and the first week's 16MP requests can get an additional 50% discount (Stand up and cheer, brothers and sisters!!) .

Testing to this point, I suddenly had a feeling —

Multimodal generation might be about to enter its own coding era.

In the past, AI-generated images often "became toys" in the commercial sector, mainly because the details couldn't stand up to scrutiny and had too much plastic feel. Who would be happy to see that as the client?

But now, H3-Image has achieved three things —

Removing AI taste, generalizable aesthetics, and logic.

When these conditions are met, the application market for multimodal technology will definitely not be inferior to programming.

From high-precision industrial packaging design, commercial illustration, to cross-border marketing materials in multiple languages, the emergence of image models can significantly shorten the traditional workflow cycle.

And when combined with the buffs mentioned earlier: speed + open-source + low cost...

What kind of future will this bring? It's hard to imagine.

But finally, let's bring our focus back to the current AI community.

When the attention of global AI companies is almost entirely focused on pure text, MiniMax still devotes a significant portion of its efforts to the multimodal world.

This has been a characteristic of this company from the beginning, betting on multimodal capabilities, and now finally experiencing emergence .

It's often said that multimodal is not important for AGI, and only coding has a business model...

But I have always thought that if the definition of AI is limited to text, this future would be too dull.

From H3, which reshaped the video generation landscape, to the current H3-Image that has made breakthroughs in logic, structure, and aesthetics.

The wild fire of domestic open source is now completely spreading into the multimodal field—

@ @ @

@ @