company

2026-11-11 00:04

In August this year, MiniMax opened the weights of the H3 video generation model, which is the first open source video model ever to reach SOTA level.

In the following months, the community played tricks based on it that far exceeded the official demonstration, and also gave birth to a number of derivative models such as H3 max, which really pushed the AI ​​creative industry a step forward.

Now, MiniMax doesn’t even intend to hide the image generation model. Just now, MiniMax H3 Image was officially released, focusing on image generation and editing, targeting SOTA in open source image models.

The previous H3 has passed unified full-modal pre-training, and has the ability to understand and generate images, videos, audios and other content, including Vincent pictures and image editing.

This time, H3 Image is based on the existing capabilities of H3 and further performs post-training optimization for image generation and editing, allowing the model to perform better in following complex instructions, image details, and visual aesthetics.

The more intuitive experience is resolution and speed. In our actual testing, it only takes 10 to 20 seconds for H3 Image to generate a 4096×4096, 4K quality high-definition picture.

This means that in the past, high-resolution images that often required waiting for tens of seconds or even longer can now be completed within a dozen seconds. 4K-level image generation is entering an era of ten seconds, and the detail advantages brought by high resolution can finally be used in more frequent daily creations.

We use H3 Image to generate a variety of highly aesthetic images, which perform well in CG, game scenes, industrial design, and various art styles.

It also focuses on strengthening the ability to follow complex instructions, supporting up to 16 reference pictures, and can understand different subject, material and style requirements at the same time.

More importantly, H3 Image will open weight on the day of release, and developers can also access it through API, or use it in the MiniMax open platform, MCode and MiniMax Conch platform.

APPSO also experienced this new model in advance. In addition to the beautiful drawings, we also want to know how complex the requirements it can understand and how detailed the image can be.

Complex requirements can be drawn accurately

In the past, when we tested image models, we often paid more attention to whether it could draw the objects that appeared in the prompt words.

But in actual work, the requirements we give designers are often much more complex. For example, how should the parts of the product be assembled, what elements need to be retained in a certain area, and what are the connections between different reference drawings. The model needs to understand these conditions before it can draw the correct picture.

This time, we started with the structural design of an industrial product and let H3 Image generate an exploded view of a futuristic coffee machine. It needs to show the structure of the bean silo, grinder, brewing module, water tank, etc., while indicating the connection relationship of each part, and maintaining the consistency of product design.

In addition to mechanical structure, word processing is also a difficult problem often encountered by image generation models. In particular, it is necessary to not only write the words correctly, but also take into account the font design, picture composition and overall visual effect.

For large-character posters, the model needs to handle text content, font styling, and visual composition at the same time. In scenes that contain large paragraphs of small text, text accuracy, information level, and typesetting will also become new problems.

We first asked H3 Image to generate a neoclassical Art Deco poster with the theme of Phoenix, which combines Western mythology, retro engravings, Victorian decorative patterns and modern minimalist typography.

H3 Image accurately renders the title PHOENIX MYTH and subtitle Eternal Rebirth of the Fire Bird.

Regarding the performance of H3 Image on small print, we found the Huang Jen-Hsun open letter that has been making a lot of noise during this period, and asked H3 Image to generate a letterhead version that was placed on the desktop.

The text in the picture is basically generated accurately, there are no obvious errors, and the overall style is consistent with the style around the picture. If you look carefully at the details around the letter paper, they are all related to Lao Huang.

We can also generate a nostalgic scene: a classic leather jacket hanging on the wall, accompanied by the monologue on the typewriter, "I've been wearing it for 30 years."

We also tried to use H3 Image to generate a more complex game interface. The text rendering and layout in the image are similar to those of survival games, showing life, magic power and physical strength levels.

Furthermore, it depends on whether Image can understand the logic between text and pictures. This requires it to have the ability to understand semantics, as well as knowledge in science, mathematics, biology, etc., and to be able to present the logic accurately and beautifully.

From mechanical structure to text layout to knowledge understanding, these cases actually point to the same question: whether the model can truly understand the information the user wants to express and accurately convert it into an image.

This time H3 Image supports up to 16 images as reference. This means that we can provide different perspectives of the product, material references, brand elements, and scene images all at once, and then ask the model to integrate this information into the same piece.

What’s more worth looking forward to is that this time H3 Image can also access Agent Harness such as MiniMax Code, Claude Code, Codex through API, and directly participate in more complex task processes.

In other words, we can directly put forward creative requirements in the Agent's dialog box, let the Agent organize the tasks, call H3 Image to generate pictures, and then continuously modify the design through subsequent conversations.

Swipe left to see more

Just like this case of using MiniMax H3 Image to generate V8 engine explosion disassembly diagrams in Codex, through Agent’s collection and organization of information, the final disassembly diagrams generated are very realistic.

For the Agent, it needs to continuously manage the context and the user's modification opinions; for the H3 Image, it needs to understand the complex instructions compiled by the Agent and execute them as accurately as possible. This way of working also puts forward higher requirements for the image model's ability to follow instructions and understand complex information.

This open source image generation model H3 Image can indeed understand complex requirements, perform precise modifications, and maintain consistency in continuous creation. This is where it begins to truly integrate into the Agent workflow.

In particular, the images generated by many current mainstream image generation models are still at one million pixels or four million pixels, while H3 Image can directly generate 4096×4096 ultra-high-definition images, with the number of pixels reaching 4K quality.

This means that in the subsequent workflow, even if we enlarge or crop parts of the picture, there will still be enough details available. Whether it is product design, commercial posters, or visual materials that need to be displayed on the big screen, there is more room for post-processing.

Just rely on a prompt word to draw a true sense of beauty

Strong understanding does not necessarily mean good aesthetics. We have seen many AIs that can strictly abide by prompt words, but the things they draw are always less interesting.

Just like GPT-Image, many people have the impression that it generates very strong information graphics, but when it comes to generating highly stylized images such as Midjourney, GPT-Image is also a bit weak.

For AI images, especially scenes such as CG, game concept design, and illustrations, many times we cannot fully describe the desired composition, color, and atmosphere. We can only give a general direction, which basically relies on the model's own aesthetic level.

This time to test the aesthetics of H3 Image, we found some reference pictures we liked on Midjourney. We first asked ChatGPT to break down the composition, light and material in the picture into prompt words, and then used these prompt words to give questions to H3 Image.

For example, in these horse racing pictures full of speed, we made it clear in the prompt words that the core style of the pictures is cinematic dynamic photography, slow-motion tracking photography and retro sports film aesthetics.

H3 Image can accurately depict the horizontal smear caused by the high-speed running of the horse, the dark green background, the horses illuminated by the golden sunset, and the large area of ​​dark, abstract negative space in the picture.

Needless to say, there is no AI flavor. When you see these photos being generated, there is even a strong sense of a blockbuster movie, after the photographers spent a lot of time retouching the photos.

We continue to use this set of prompt words to test different sports scenes, from racing, cycling, skateboarding, to moving cars, running people and animals, to see if H3 Image can present the beauty of speed and the imperfect but moving texture in photography.

Each picture is very emotional, with the moving subject gradually dissolving into traces of light, color and time.

For the generation of portraits, we asked H3 Image to generate several pictures of retro direct flash film photography + lifestyle documentary photography + independent fashion magazine aesthetics.

The resulting picture is like a random moment captured by a friend at a private party. There are no carefully arranged poses, but it is more vital than a formal photo.

Of particular note is the realism of the characters shown in these images. The skin retains its natural texture, and the details of hair, clothing, and light and shadow are rich enough, without the over-skinned and greasy texture that is common in many AI-generated portraits.

When copying this style to other images, you can also see H3 Image’s ability to control the style. The strong light and dark contrast brought by the direct flash at night, the uncontrollable moment when the character laughs naturally, the slightly messy composition, the rich floral colors, and the unmodified texture of the film snapshot are all successfully generated.

Similar portrait photos include a movie still with oriental classical aesthetics that we generated using H3 Image, integrating Chinese costume film photography, naturalistic light and shadow, and retro film visual texture.

These pictures seem to recreate an ordinary and intimate moment in ancient life. The afternoon sunlight passing through the leaves gently illuminates the characters' faces. The natural expressions and mottled tree shadows turn the picture into a movie memory that will be treasured by time.

In the process of reproducing the actual game screen of GTA V, we also specified the style and third-person perspective in the prompt word, and asked H3 Image to generate an open world 3D game screenshot with the actual visual style of the "Grand Theft Auto V" game.

H3 Image fully understands that what we want is "GTA game graphics", not a simple "GTA style". Several pictures are like a frame randomly captured by players in the GTA open world. They are different from other AI-generated pictures. Other AI-generated pictures are more like official carefully produced promotional pictures.

We also generated several highly stylized pictures, all of which are in the form of Vincentian pictures generated directly at one time. There is no additional picture reference. With only one-time prompt words, H3 Image can be aesthetically completely AI-free while also accurately depicting the style we require.

For creators who already have reference images, we can use GPT-Image to extract the corresponding style prompt words, or we can directly use the reference function of H3 Image to apply the focus tracking method of a certain photo, or an illustration to expose the brushstrokes on the bottom of the paper, and then put it on another subject.

When complex instruction understanding and visual aesthetics can be taken into consideration at the same time, the image model will have more practical uses. From game concept drawings, brand posters, to product structure diagrams and design plans, the number of times users need to repeatedly explain their needs and adjust prompt words has the opportunity to be further reduced.

From video to images, MiniMax is opening up all-modal capabilities

Back to H3 Image, the most noteworthy aspect of this release should be open source.

Whether it was the video generation model at the beginning or the current image generation, these are areas that have long required high computing power, training data and model capabilities.

Especially the top visual generation models, in many cases, we can only use them through official products or APIs. What the model can do, how it operates, and how it can be further modified still depends on the service provider.

MiniMax's open source has transformed the model from a company's product into a starting point for more users and developers to continue to create: they can not only use ready-made models, but also have the opportunity to deploy, fine-tune and optimize according to their own needs, develop new creation tools around it, and truly integrate model capabilities into existing products and workflows.

For content and the wider industry, H3 Image, as a SOTA open source image model, gives everyone a new choice. For example, game developers can try to establish a material generation process that matches their own art style, design teams can establish automated workflows for specific products, and individual developers also have the opportunity to build their own creative applications around the open model.

From videos to images, combined with text, voice and other models, MiniMax's multi-modal layout is gradually forming a more complete capability matrix. H3 has proven once that when the weight is opened, the gameplay, derivative models and workflows emerging from the community far exceed the boundaries of the official demonstration, and it has also verified that the open source model can also support solid commercialization.

For us, the most exciting thing is that the first-tier visual generation capabilities can no longer rely on a few closed services.

When generative models that are smart enough and good-looking enough can be made available to developers, more people will have the opportunity to turn their ideas into tools and works that truly belong to them.

*All images in this article were generated using MiniMax H3 Image

Zhang Zihao

A total of 201 articles have been published

This article comes from the column

Previous article2026-11-10 17:30

Ternus puts pressure on Siri boss, Apple AI can’t wait any longer

Aifaner App

Love style, make the future within reach

Digital trendy brands focusing on tomorrow’s products

Advanced tools, prophets first, AIGC’s inspiration guide

A new era of car manufacturing, tomorrow’s travelers

Explore new lifestyles in the city and be your guide to tomorrow’s life

Copyright © Guangzhou Liwo Zhiyuan Investment Management Partnership (Limited Partnership) 2008 - 2026. Use of Aifaner website content for commercial purposes requires permission. Non-commercial use is licensed under CC BY-NC 4.0.

All content is made available under the CC BY-NC 4.0 for non-commercial use. Commercial use of this content is prohibited without explicit permission.