MiniMax Open-Sources H3 Image: Native 4K in 10 Seconds, Up to 16 Reference Images

10 Seconds to 4K: MiniMax Open-Sources H3 Image!

Author: 老金带你玩AI

MiniMax has open-sourced its 4K image generation model, named H3 Image, which now ranks at the top of the open-source SOTA list.

I (Lao Jin) stayed up all night playing with it...

According to the official speed, it generates a 4096×4096 image, about 16 megapixels natively, in under 10 seconds (I got about 1 minute on my 5090).

The weights are already released, and ComfyUI has ready-made templates, so you can do whatever you want.

Let's look at the results first; I'll show you how to use it in ComfyUI later.

Here are some examples from the internet, then we'll look at my ComfyUI test.

First, this map. Point to a location on the map and let AI draw the scenery you'd see standing there. What would it look like?

Stand at the red arrow and look in that direction.

!!!

If the image is compressed, it's the website's compression; I saw it as ultra-high-definition, quite stunning. Ultra-high-definition is really large.

A suspension bridge crosses the river, with a waterfall and a castle on the mountain behind. A moment ago we were looking at the map, and now we can look along the river. I felt like pressing W (the forward key in games).

This is MiniMax's official demo: give it a map, a location, and a direction, and it generates the corresponding game-style scene. When drawing novel scenes or game storyboards, this kind of labeling is much easier to understand than repeatedly describing "a bit left, a bit farther".

Now let's add furnishings to a moon base.

After the riverside, let's go to the moon.

In this set of official reference images, the vehicle, capsule, and astronaut are arranged, looking like white models (untextured meshes, also called gray models).

Let H3 Image keep the layout and fill in the materials and lighting.

The car body now has shading and reflections, the tires show rubber texture, and the ground dust has light and shadow differences. Everything else remains unchanged; the sci-fi atmosphere comes out.

If you have a white model with layout already set, you can follow this idea: explicitly describe metal, glass, dust, and light.

This old outfit has a lot of detail.

We can fill in the scene materials; how do we describe the character's texture?

Community user @AIWarper made this one. Don't look at the face; let's follow me and look at the collar.

Stitching, wear, and collar folds are all there, and the gold makeup on the face is cracked. The skin keeps its texture, the clothes show old marks, and these details together look incredibly real.

The author noted it is 6272×2688, a 21:9 widescreen ratio. For cinematic character portraits, besides saying "cinematic", it's best to specify makeup, fabric, lighting, and other concrete elements.

Eight people, eight color schemes, and a group photo.

A single close-up can pile details on one person; @Just_sharon7's set instead drops in eight people. Each has a specified color scheme, with the color swatches in the lower right of the reference image corresponding to the character numbers.

Note them: for example, the second person is pink, the seventh is purple.

Now look at H3 Image's group photo.

The person numbered 2 in the original image now wears a pink shawl and wine-red dress; the person numbered 7 is all in purple. The clothing shapes, standing positions, and sitting postures of all eight are differentiated. It's a unified group photo while respecting each person's outfit preferences.

This set uses a reference collage, incorporating character photos and color cards. When replicating, number the characters and palettes, then specify each person's outfit and position.

Below is the author's original prompt in Chinese, partially translated:

Use characters 1 to 8 from the reference collage as eight roles, preserving each face, hairstyle, and body proportions. Assign the eight color schemes from part 9 to the eight people.

Primary colors determine the main body of the outfit; secondary colors are for seams, layers, and structural zones; accent colors are for accessories, shoes, and local materials. Redesign the clothing's silhouette, cut, fabric, and structure.

In a neutral gray studio, arrange the group through standing, sitting, leaning, front-back distance, and height differences. Give each color scheme an independent visual zone.

H3 Image supports up to 16 reference images.

Characters, clothing, and color schemes can also be uploaded separately, and you can use image numbers to describe correspondences.

Mech full-view.

That green mech standing there already has shoulder armor, weapons, and leg armor.

For character design, you also need to know what it looks like from behind. @Ankit_patel211 had H3 Image take this image and produce a multi-view line art sheet.

For this kind of setup, upload the character image, request front, back, left, and right views, and ask for black line art on white, with all views aligned at equal height.

Beyond character design, @iam_mian7 also put a character and a game UI in one image.

The face on the right is almost touching the camera, with water droplets clearly visible; on the left are spaces for class, level, skills, and equipment.

Furniture arrangement is also a matter of minutes.

Sometimes furniture is bought piece by piece and doesn't match. You might like the red tea table, the green chair, and the striped footstool. But when they all squeeze into the living room, will they still look good?

The official gave H3 Image an empty room and nine pieces of furniture, and it arranged them like this.

First let the furniture meet in the image, and if it works, consider moving it home. Just give a room image, swap whatever you want, and describe positions clearly.

Chinese text, laid out with explanation.

Those images were light; this last one throws a problem at us. The official gave a reasoning puzzle about personnel selection: five conditions on top, selected and unselected people in the middle, and reasoning steps below.

I came to look at image generation, but here my brain also had to go to work.

Where are the weights, and how to use ComfyUI.

After seeing the images, if you want to try it, I'll show you how to get it onto your own computer. The cloud version on Hailuo AI is paid; let's go straight to local deployment.

For ComfyUI, just open it and download here. ComfyUI needs 0.45.0 or newer, and the template pack needs 0.11.104 or newer.

First let an agent (CC or Codex or whatever) find a quantized version that fits your machine. Don't blindly use the official one or you might blow up your memory.

If your VRAM and RAM are enough, you can use the official version; find H3 Image here.

Open it, take a look, then have your agent read it and either copy a config that matches your local models or download the one it has prepared.

I tested some samples; here are a few.

For example, a high-precision mechanical exploded view.

And changing my outfit.

There's also this 16-image fusion (requires luck, but success rate is quite high).

The effect is quite good.

And this multi-Chinese text (MiniMax's long-standing pain; it still breaks a few characters, so retries are needed).

Finally.

Native 4K large images, sub-10-second generation, and support for up to 16 reference images: these three together are quite compelling. You can zoom into character details, dress eight people using their own palettes, and even turn a storefront visual into web pages and menus. High resolution and multi-reference really pay off in these kinds of tasks.

Even better, MiniMax has opened the weights, and ready-made ComfyUI templates are available. After seeing the examples, if you want to try it yourself, the models and instructions are all prepared.

This open-source image generator is indeed a formidable player.

Thank you for reading my article. If you found it useful, please give it a like, share, and forward. If you want to be notified first, star me. Thanks for reading.

Open-source knowledge base (real-time updates, discussion group): https://my.feishu.cn/wiki/OhQ8wqntFihcI1kWVDlcNdpznfF