MiniMax H3 Image Edit

MiniMax H3 Image Edit — AI Image Editing (Image-to-Image)

What is MiniMax H3 Image Edit?

MiniMax H3 Image Edit is MiniMax's instruction-driven image editing model, built on the open-weight H3-Image checkpoint. It takes one to sixteen reference images and a single text instruction and returns an edited image — combining subjects, swapping or restyling elements, and composing new scenes while keeping the parts you tell it to preserve. You name each input in the prompt as "Picture 1", "Picture 2", so the model knows exactly which object to move, change, or keep. Output is native high resolution — up to about 16 megapixels (4K) in a single pass, with no separate upscaling step.

Key Features

• •Edit and combine 1–16 reference images from one text instruction.
• •Native output up to ~16 MP (4K); choose a 1K, 2K or 4K area.
• •Any aspect ratio from 1:3 to 3:1, or an exact WIDTHxHEIGHT size.
• •Strong instruction following — renders exactly what the prompt asks while preserving the parts you name.
• •Standard and highspeed tiers; 1–10 images per request; JPEG, PNG or WEBP output.
• •Returns the enhanced prompt (revised_prompt) the model actually read.

Best Use Cases

Product compositing (drop a product into a new scene, or combine several products), virtual staging and set dressing, brand-consistent variations from reference art, character and prop combination, outfit and material swaps, and background replacement. Because the model holds identity across a large change, it fits catalog, marketing and concept work where the subject must stay recognizable. In testing, two separate toy photographs were combined into one coherent 4K shelf scene with each object's texture and markings carried over faithfully.

Prompt Tips and Output Quality

State both the change and what must stay the same, and refer to inputs as Picture 1, Picture 2 in upload order. Keep reference shots clean and well lit. Use area=4K for print-grade detail, 2K for faster iterations. Set an explicit aspect_ratio (for example 4:3) for predictable framing, or auto to let the model infer shape from the references. Prompt enhancement runs on every edit, and the rewritten prompt comes back so you can see what was generated.

How to Use MiniMax H3 Image Edit

MiniMax H3 Image Edit turns one to sixteen reference images plus a text instruction into a single edited image. The model is tuned for instruction following, so the clearer you are about what to change and what to keep, the closer the result lands.

Prepare your references

Upload clean, well-lit shots of each subject (JPG, PNG, WEBP or HEIC, 128–8192 px a side). In the prompt, name them in upload order — Picture 1, Picture 2, and so on. This tells the model which object to move, combine, or preserve.

Write the instruction

Describe both the edit and the constraints: "Place the bottle from Picture 1 on the shelf in Picture 2; keep the label and the shelf unchanged." Spell out the textures, colours and markings you want held. Prompt enhancement runs automatically on edits, and the rewritten prompt is returned as revised_prompt, so you can see exactly what the model read.

Recommended Settings

• •area: 4K for print-grade detail (native ~16 MP); 2K for faster, lighter iterations; 1K for quick drafts.
• •speed: standard for the best quality; highspeed when turnaround matters more.
• •aspect_ratio: auto to infer shape from the references, a fixed ratio such as 4:3 or 16:9 for predictable framing, or an exact size (WIDTHxHEIGHT).
• •num_images: 1 returns the image itself; request up to 10 to compare variations.
• •output_format: jpeg for photographs, png for lossless, webp for small files.

Tips

Combine several products or subjects in a single pass by referencing each Picture explicitly in the instruction. For brand-consistent variations, reuse the same reference set and change only the instruction. Keep the number of simultaneous changes modest for the cleanest identity preservation — in a tested two-image combine, each object's texture and markings carried over cleanly at 4K.

Examples to try

Each example ran end to end in the playground at the settings shown, from reference images named Picture 1, Picture 2 in upload order — the model takes one to sixteen.

• •Two-reference combine, identity preserved. Prompt: "Picture 1 is a mustard hand-knitted plush elephant; Picture 2 is a wooden pull-along giraffe. Seat the knitted elephant from Picture 1 upright on the back of the wooden giraffe from Picture 2, on a pale birch nursery shelf with a window of diffused daylight to the left. Keep the elephant's chunky knit stitches, its red felt heart and black bead eyes exactly as in Picture 1, and keep the giraffe's painted brown spots, red pull-string and black wheels exactly as in Picture 2." Both toys land in one 4K shelf scene with every stitch and marking carried across. Tip: name the textures to hold, not just the objects.
• •Restyle / medium transfer. Prompt: "Picture 1 is a photograph of a flamenco guitarist seated on a stool on a small wooden stage. Repaint Picture 1 as a loose oil painting with thick visible brush strokes and a warm earth-tone palette, keeping the guitarist's pose, the guitar and the wooden stage the same." The photograph returns as a thick-brushstroke oil painting, pose and layout intact. Tip: say what to keep so only the look changes.
• •Aspect-ratio control. Prompt: "Picture 1 is a tan leather lounge chair; Picture 2 is an arched brass floor lamp. Place the lounge chair from Picture 1 in the corner of a concrete-floor loft beside a tall steel-framed window, and stand the arched brass floor lamp from Picture 2 behind it so the lamp curves over the chair. Keep the chair's tan leather and stitched seams exactly as in Picture 1, and the lamp's brass arc and round shade exactly as in Picture 2." Both pieces share one loft corner, framed tall at 9:16 with neither one cropped. Tip: set aspect_ratio for a fixed shape, or auto to let the references decide.

Further instructions this model is built for: "Swap the jacket in Picture 1 for the denim one in Picture 2, keep the pose", "Change the wall behind Picture 1 to exposed brick, keep the furniture unchanged".

How many images can I edit at once?

Up to 16 reference images per request.

What resolution does it produce?

Up to about 16–20 megapixels natively, with no upscaler.

How do I point the model at a specific input?

Name it Picture 1, Picture 2 in the prompt, matching upload order.

Which file formats are accepted?

JPG, PNG, WEBP or HEIC inputs; JPEG, PNG or WEBP output.

Can I get several variations in one call?

Yes — request up to 10 images per request.

Does it return the enhanced prompt?

Yes, as revised_prompt in the response.