prompt: This is a realistic vertical photograph. A row of silver stainless steel subway carriages with signs of aging is parked in a rough underground station. The side of the carriages is covered with a large wild-style graffiti that spans the entire row of windows. The graffiti consists of sharp-edged 3D block letters spelling out "MiniMax H3 Image"; the letters are filled with bright orange and have gradient effects, accompanied by deep blue 3D shadows, thick black outlines, and thin white highlights. Even fresh spray paint drips can be seen. In the background, the industrial-style ceiling is exposed, revealing rusty steel beams, pipes, cables, and cool-toned fluorescent lights. The concrete walls are covered with stains and dirt, and below them are the tracks and gravel ballast. The image presents a cool fluorescent lighting effect and metallic reflective texture, processed with cinematic color grading, featuring fine film grain and slight vignetting, creating an authentic New York subway atmosphere. The whole photo seems to have been shot with a 35mm lens, rich in details, showcasing a raw and authentic urban street photography style.
The long string of letters "MiniMax H3 Image" not only has the complete spelling but also features 3D modeling, gradient fill, and spray paint effects. Text generation capabilities that can achieve this already have a hint of visual design.
Fashion portrait wants a magazine look , it's not just about looks; facial features, clothing details, and lighting atmosphere all need to be on point. We challenged a set of Victoria's Secret backstage shoot:
prompt: Create a glamorous photoshoot in the style of Victoria's Secret. A young woman attached in the uploaded reference image (Keep the person's face 100% accurate from the reference image) stands almost sideways, slightly bent forward, during the final preparation for the show. Makeup artists apply lipstick to her (only her hands are visible in the frame). She is wearing a corset decorated with beaded embroidery and crystals, along with a short fluffy skirt and large feather wings. The image has a "backstage" effect. The background is a darkly lit room, probably under the podium. The main focus is on the girl's face and the details of her costume. Emphasize the expressiveness of her gaze and the luxurious look of the outfit. The photo is lit by a flash from the camera, highlighting the shine of the beads and crystals on the corset as well as the girl's glossy skin. Victoria's Secret style: sensuality, luxury, glamour. Very detailed. Important: do not change the face.
The intricate details of complex elements like crystal beads and feather wings are fully rendered. With direct flash lighting, the model's skin glow and the outfit's reflective layers are distinct. A single shot could be a page from a fashion magazine.
Try a different style, let's try movie-style portrait:
This street portrait has a strong movie screenshot feel, with the moment of the subject looking back clearly captured, while the surrounding pedestrians are processed into large dynamic motion blur. The contrast between still and moving elements creates immediate visual tension, and the overall atmosphere, sense of cinematography, and emotional expression are all on point.
We have also unlocked the retro 2000s photo style.
Prompt: Without changing her original face, create a portrait of a beautiful young woman with porcelain-white skin, captured with a 1990s-style camera using a direct front flash. Her messy dark brown hair is tied up, posing with a calm yet playful smile. She wears a modern oversized cream sweater. The background is a dark white wall covered with aesthetic magazine posters and stickers, evoking a cozy bedroom or personal room atmosphere under dim lighting. The 35mm lens flash creates a nostalgic glow.
The slightly grainy film texture, the old magazine posters on the wall, and the warm yellow string lights instantly transport you back to the 2000s. The slight overexposure also gives a vintage digital camera feel. The atmospheric details in the prompt are all accurately recreated.
After generating from text, let's try image-to-image. This part requires the AI model to understand the original image, then modify it according to the instructions .
Start with Elon Musk. Upload a photo and let the model turn him into a cute and funny hand-drawn WeChat emoticon. The original expression of the subject is exaggerated and magnified, paired with wobbly "So Dumb", sweat drops, and question marks, resulting in an emoticon ready to be used for playful arguments.
Prompt: Draw him as a funny hand-drawn WeChat emoticon. Use a simple, cute and clumsy line drawing style, with a bit of the feel of a casual doodle. The background is pure white. According to the expression in the original photo, exaggerate his shocked, disgusted, or lazy look. Add a few sweat drops, question marks, or little flashing stars around his head as cute doodles. Add handwritten English "So Dumb" at the bottom, with a wobbly and messy handwriting style, making it look casual and humorous.
The next round increases the difficulty. Upload a street CCTV surveillance photo, asking the model to identify the people in the image and add white face bounding boxes around each person, then generate a high-resolution close-up of the most prominent figure.
The generated result preserves the original scene layout and surveillance perspective, maintaining the appropriate grain and slight distortion in the main image, while the faces within the magnification boxes are clearer. The overall effect resembles a generated analysis image from a surveillance system.
Prompt: Create a high-angle CCTV surveillance shot using the uploaded image as the source. Detect every visible person in the image and automatically draw a white rectangular bounding box around each face. For the most prominent person, add a large zoom in inset: a sharp, enhanced close-up of their face displayed in a floating rectangular frame connected with a thin white line. Keep the main image slightly noisy and security camera like (soft grain, slight distortion, muted colors), while the zoom in face box should be clearer, brighter, and more detailed. No text, no timestamps, no overlays except the boxes and connecting line. Maintain the original scene layout, angle, and environment of the uploaded image.
If you want to performstyle transfer, such a realistic style image can be transformed into a pastel style collage:
the effect also looks quite good.
It is evident that MiniMax H3 Image can not only generate photo-realistic images but also help you realize some creative ideas, and even let you participate in image battles. This time, the effect seems to have improved compared to before,realism is fully loaded, and the AI taste has become very subtle.It's also worth mentioning itsinstruction following ability. When facing complex information, this model can well understand and output appropriate results.
This makes us look forward to its work efficiency.
Understanding of the physical world and instruction following
Most of what we tested above is about the realism of AI-generated images. Now we will examine something more challenging:does the model really 'understand' what it is drawing? This is very important for some practical design work.
In this aspect, the model needs to have a correct understanding of the structure and materials of the physical world, organize information clearly when there are dense information points, and accurately follow instructions for adjustments.
Let's try a high difficulty one first: have MiniMax H3 Image generate an exploded view of an industrial blueprint. The key is not only aesthetics, but the model must also have knowledge of physical structures. It needs to know what parts make up the camera, how they are assembled and nested, and where each part is located after disassembly. Previously, AI often failed in this regard.
Prompt: Please generate a poster of a retro-style technical information illustration. An exploded view of a classic 35mm rangefinder camera, with a 16:9 horizontal layout, a warm cream-colored background. Seven camera components are vertically suspended along the central axis, precisely aligned, and spaced between layers: the top cover with black leather with a pearlescent texture and a metal focusing knob, a small viewfinder, a silver lens tube engraved with distance markings, four glass lens elements and a brass aperture mechanism, a brass shutter gear assembly, a dark film transport gear set and crank, and a metal base plate. Each component is connected to a black circular numbered label with fine lines (white numbers 1–7). Top left corner: a large serif font title "VINTAGE 35mm CAMERA", subtitle "Exploded Technical View", a thin separator line and small explanatory text. Two right-side thin-lined border panels: "PARTS LIST" with seven numbered entries (bold names + gray explanatory text), and "SPECIFICATIONS" listing film format, lens, shutter speed, and weight. A detailed technical illustration style, delicate pen sketch combined with soft realistic rendering, realistic textures of brushed metal, polished brass, transparent glass, and leather, soft shadows, low saturation retro color palette, clean engineering manual aesthetics, flat information graphic composition, high resolution.
It is evident that the generated image shows the body, lens, gears, and other parts being disassembled layer by layer, with numbered labels, English explanations, and parameter tables also generated. The overall quality is almost at the level of product manual illustrations.
Or upload a flat layout image, and H3 Image can generate highly accurate photos of real people wearing the outfit, stealing the livelihood of fashion bloggers.
Prompt: show woman wearing the outfit.
E-commerce is one of the most practical applications for AI image generation.
Nowadays, many merchants have started using AI to create product main images, detail pages, and marketing posters. However, the real test for the model is whether it can balance product texture, brand information, and visual composition, ensuring that different elements do not clash on the same image.
The materials generated by H3 Image highlight the product clearly, with textures, packaging, and details that hold up under magnification. The surrounding decorative elements also align with the product's selling points. In terms of multi-layered text layout, the brand title, product name, and promotional information form a clear visual hierarchy.
In terms of flat design, we want not only aesthetic appeal but also accurate text on the poster. Can putting dozens of information points into one prompt cause confusion? We asked the model to create a poster themed around oriental camellia, while also handling the main title font design, Chinese and English text, floral composition, and text layout. The phrase "Shanjian Huaxu" (Mountain Flowers in Sequence) is interwoven and merged with petals, with clear textual hierarchy:
Prompt: extreme minimalism and classical aesthetics, high-end poster layout, oriental camellia art, detailed petal textures, white to light pink gradient effect, artistic concept style, visually elegant and rich in Eastern classical atmosphere; the main title is placed in the lower right corner, designed with an extremely thin Songti style font "Shanjian Huaxu," arranged in a staggered and interwoven manner, with elegant and elongated characters, floral curves added at the ends of strokes, naturally interwoven among each other, creating an overall Eastern poetic feel, exploring the artistic expression of natural flowers and the micro world of plants; small text at the top of the image sequentially displays "2026," "BOTANICAL ART," and "2026," while the text on the lower left emphasizes the time numbers "TIME——08.18/08.24," and the image incorporates small English text "Camellia in Bloom," "Oriental Botanical Aesthetics," and "A Poetic Expression of Nature."
Finally, let's check H3 Image's ability to follow instructions . For many people, the most commonly used scenario in daily work is image editing, with adding, deleting, modifying, and replacing being almost basic operations, all of which are clear semantic actions.
Thinking about it, today Elon Musk seemed to express "honor" for having the "opportunity to hire Jensen Huang." Let's imagine this scene: Jensen Huang holding Sink receives a warm welcome from Elon Musk at the SpaceX gate, generating this image requires a series of instructions and several reference images:
He quickly joined in.
After a round of testing, MiniMax H3 Image is a model that excels in all aspects. In terms of aesthetics, H3 Image can consistently produce photo-realistic images, maintaining facial features, style transformation, and fine details. In terms of productivity, the completion of text layout, local editing, e-commerce posters, and industrial schematic diagrams is sufficient to support workflow . From daily use to professional production, open-source AI image models have reached a usable level.
After reviewing the practical test results, we began to wonder what kind of model MiniMax H3 Image really is.
How did the video model become the SOTA in image generation?
Unlike most image generation models on the market, H3 Image is derived from MiniMax's full-modal video model H3 , which was released by MiniMax in July and made its weights open. This system uniformly understands text, images, videos, and audio within a single architecture and can generate videos with native stereo sound.
H3 Image adopts the overall design of H3-Base, removing the audio branch and retaining the backbone network, text encoder, and other architectures. On the text side, it directly uses the text encoder from H3, a multimodal language model with 32 billion parameters, responsible for understanding complex instructions, long descriptions, and multi-condition logic. On the image side, the reference image will enter two modules at the same time: the text encoder is responsible for understanding "what's in this image," while the image VAE is responsible for compressing pixels into visual latent variables that the generative model can process.
The core backbone of H3 Image is: a diffusion Transformer (DiT) with 331 billion parameters, which adopts the same three-dimensional multimodal rotary position encoding as the H3 Omni Transformer. Text tokens and reference image tokens are packed into a complete sequence and fed into this network in a single stream, directly predicting the latent variables of the target image, which are then decoded into images by the VAE.
In terms of resolution, H3 Image chooses native output, without any external super-resolution modules. A single checkpoint can generate images with approximately 16 million pixels, covering various aspect ratios from 4096×4096 to 6272×2688 (21:9).
H3 Image also has a fast generation speed: the median P50 time for 4K image generation is about 10.3 seconds for the fast version and about 16 seconds for the standard version, while most comparable models in the same category take over 50 seconds.
MiniMax's approach of extending from video models to image models is the opposite of the traditional graphics-first approach. In fact, this has many advantages for image generation, as video models need to learn about different perspectives of objects, changes in material under lighting, and movement patterns of people and objects to maintain consistency across continuous frames.
MiniMax H3 already has rich visual world knowledge, and H3 Image, based on it, naturally produces good results.
The global AI ecosystem is increasingly relying on domestic open-source technologies.
As AI technology develops faster and faster, the sudden increase in the capabilities of new models seems not always to require "shock and awe." Actually, what has left us in awe is the excitement of the global AI community around the release of MiniMax H3 Image.
Now that MiniMax has released a new model, it's already a big deal for the AI community, and this enthusiasm has already transformed into the flourishing of the ecosystem. MiniMax is gradually becoming a foundational platform for other AI startups to conduct technical development.
In fact, before H3 Image, an overseas model called P-H3-Image appeared.
This is an open-source image generation model, and a startup company named Pruna also adapted MiniMax-H3 into a high-performance image model, not only achieving high-quality image generation but also having very low inference costs, aligning with the official's vision.
Now that H3 Image has been released, "the official version has arrived."
In fact, it's no longer rare to further pre-train, fine-tune, or perform reinforcement learning on the weights of H3 Image to release new models with improved capabilities. On the day M3.1 Flash opened its weights, there was already an American startup that released a specialized model adapted for it, helping M3.1 Flash generate text more quickly.
The derivative ecosystem around MiniMax H3 is more active. The NVIDIA team released Sol-H3, optimizing the complete inference pipeline for video generation; Pika released Pika H3 Flex, reducing the number of generation steps and improving the image quality and creative control. The model tree based on MiniMax-H3 on Hugging Face has already collected about 220 derivative entries, with some startups transforming H3 into independent image generation models, video character replacement models, and even interactive world models. The business model is gradually taking shape within this ecosystem. By offering free weights to occupy the niche, MiniMax is using hosted APIs and subscription plans to generate revenue from usage. The revenue thresholds and hosting limitations in the community license will guide large-scale players towards commercial licenses. Long-tail free and scale-based charging enable open source to build its own commercial loop. Looking back at this week, the capability boundaries of image generation models are redrawn every few days, but what truly determines the long-term landscape may no longer be about higher scores. The open source ecosystem is defining the breadth of capability diffusion. When 4K generation and dozens of reference images become accessible to everyone, the core of competition has quietly shifted from "whose model is stronger" to "whose ecosystem is better." The next breakthrough in AI generation might emerge from a developer who downloads H3 Image weights.
