MiniMax H3 Image prompt guide

Want real prompts to copy? The H3 Image prompt library is an image-first gallery of real prompts from public X posts and official MiniMax examples. Text-to-image cards show pictures we generated ourselves from those prompts; authors are credited with a link to the original post.

A candid smartphone photo of [person: age, look, clothing] [doing something] in [place], natural window light from the left, slight motion blur on the hands, realistic skin texture with pores and small imperfections, shallow depth of field, everyday colors, shot at eye level from about two meters away, vertical 3:4.

Character reference sheet of [character: name, age, build, hair, outfit, signature item] on a plain light-gray background. Three full-body views in one row: front, three-quarter and back, same scale and pose, plus four head close-ups with different expressions (neutral, smiling, angry, surprised) below. Consistent colors and proportions across every view, clean even studio lighting, thin labels under each view reading "FRONT", "3/4", "BACK". Landscape 16:9.

Exploded-view technical illustration of a [product], every component separated along its assembly axis and floating in order with even spacing, thin leader lines to short labels for each part ([part 1], [part 2], [part 3]), three-quarter isometric angle, matte white background, soft shadows, precise industrial-design rendering, crisp legible label text. Landscape 16:9, native 4K.

Top-down knolling flat-lay of [theme] essentials: [item 1], [item 2], [item 3], [item 4], [item 5], arranged on a [surface color/material] in a tight grid with equal gaps and right angles, nothing overlapping, soft diffused overhead light, true-to-life materials and colors, commercial catalog photography, square 1:1.

A minimalist event poster for "[EVENT NAME]". Headline "[EVENT NAME]" in large bold condensed sans-serif across the top; below it "[DATE] · [CITY]" in smaller letters; at the bottom "[ONE-LINE TAGLINE]". Center image: [one simple striking object] on a flat [color] background. Clean grid layout, generous margins, high contrast, every word spelled exactly as written, no other text. Vertical 2:3.

Image 1 is the main scene. Place the person from image 2 in the center of image 1, wearing the jacket from image 3 and holding the product from image 4. Keep the person's face, hairstyle and body shape from image 2, the jacket's color and logo from image 3, and the product's shape and label from image 4. Match the lighting and perspective of image 1 and keep everything else in image 1 unchanged.

In image 1, change only [the thing to change] to [new version]. Keep the composition, people, faces, lighting, colors and all other objects exactly as they are.

MiniMax-H3-Image is built to read a detailed English description of the whole image. Everything on this page comes from MiniMax's and ComfyUI's documentation; we show no generated images of our own.

1. Short prompt in, long caption out

On the API, text-to-image runs prompt enhancement (H3-Image-Context-IR) by default. You can write a short prompt in any language and the service expands it. The exact prompt used comes back in data[].revised_prompt.

Official example: the short instruction

A rustic, overhead shot of a farm-to-table feast, showcasing a variety of colorful, seasonal vegetables, artisanal cheeses, and freshly baked bread on a weathered wooden table.

was expanded to a caption that begins:

The image is an overhead, top-down photograph of a rustic farm-to-table feast spread across a weathered, dark brown wooden table with visible grain, knots, and a worn finish. In the center, a large woven basket holds a freshly baked rustic loaf of bread…

Text from MiniMax's official examples folder (instruction.t2i.txt, revised_prompt.t2i.txt). Enhancement also picked a 4736×3552 (4:3) size for it.

2. What a good caption covers

The expanded caption above shows the pattern the model expects. Notice what it spells out:

• Shot type and medium: "an overhead, top-down photograph".
• Layout by position: center, left, right, lower left, upper right.
• Materials and textures: wood grain, cracked crust, rind.
• Lighting direction and mood: "natural and warm, coming from the upper left and casting soft, long shadows".

A reusable skeleton (our wording, untested; adapt freely):

The image is a [medium: photograph / 3D render / watercolor] of [main subject], seen [camera angle].
In the center, [subject details: pose, clothing, materials].
To the left, [...]. To the right, [...]. In the background, [...].
Any text in the image reads "[exact words]" on [surface].
The lighting is [source, direction, color], creating [mood].
[Things to avoid, phrased positively: "plain empty sky", "clean unmarked walls".]

Launch-day example of a long structured prompt (the author says a whole JSON prompt fit in the 32,000-character limit):

3. Fine-tune with revised_prompt

MiniMax's guide: take revised_prompt, edit it, then send it as the prompt of a new text-to-image request with disable_prompt_enhanced: true. When enhancement is off you must set size or a fixed aspect_ratio, and only about the first 4,096 tokens are read.

4. No negative prompt

Guidance is fixed at 1, so negative_prompt is rejected. Describe the result you want, including what should be absent, in the main prompt.

5. Editing with references

Reference order sets the numbering. API examples refer to "image 1", "image 2"; the local and ComfyUI pipelines label them <Picture 1>, <Picture 2>, up to 16. The first picture is the one being edited.

Replace the cat in image 1 with the Shiba Inu in image 2, keep the background unchanged

Official edit example instruction:

Take the over-ear headphones off the man on the left and fill in his hair where the band flattened it. Have both hosts lean toward each other and fist-bump directly above the "MONEY TALKS" desk sign…

Its enhanced version first names every element as a subject (for example "The wooden desk sign with raised lettering reading 'MONEY TALKS' in <Picture 1> is <Subject 4>"), then describes the target image and what is preserved. Edits on the API always run enhancement.

6. Text in images

MiniMax's showcase includes signs, a tarot card title and logic-puzzle boards in Chinese and English. Put the exact words in quotes and say where they appear, as the official captions do.

7. A photoreal tip from early access

Early-access tester Brent Lynch suggests the word "snapshot" to push photoreal output:

Size matters

The model works best around 16 MP and effectively from about 4 MP to 20 MP. Pick one of the native sizes on the ComfyUI page or the API's area: "4K".

8. A workflow from a hands-on review

Chinese tech outlet APPSO described this approach in its hands-on: take reference images you like, have a chat model break down composition, lighting and materials into a prompt, then give that prompt to H3 Image. For a style you already have as an image, the alternative is to pass it as a reference image instead. APPSO also reports calling H3 Image through the API from agent tools such as MiniMax Code, Claude Code and Codex.

More real outputs: MiniMax H3 Image examples.