How to prompt MiniMax H3 Image

Write the prompt as one flowing description of the finished picture, in English, in roughly this order: what kind of image it is and where we look from; the main subject with its visible details; the setting; every other element with its place in the frame; any text; the light; then the colors and the finish. MiniMax's own example prompts run to a few hundred words, and a description that long is read whole.

Content comes first, style after. Decide what is in the picture and where, then say how it looks. Replace mood words with what makes the mood: not "cozy" but a low lamp, a wool blanket, and rain on the window.

For anything made to be read, such as a poster, a menu, a product sheet, an infographic, or a screen, plan the layout before you write. Split the canvas into areas in reading order (title, subtitle, main picture, side panel, row of cards, footer) and give each one its position, its exact words, and what it shows.

Name what each area really contains. "Icons" or "some details" gives the model nothing to draw; "three round icons for sun, water, and soil, each with a one-word label below" does.

MiniMax H3 Image renders text well, from one word on a mug to signs and posters with several lines. Put every word in double quotes, exactly as it should read, and keep what you say about it outside the quotes: a sign that reads "OPEN", in red neon script.

For text on several lines, quote each line on its own and say how the lines stack: a shop window painted with three lines, "FRESH", "FLOWERS", and "DAILY", one above the other in white brush script. Keep the words in their own language and script inside the quotes; English and Chinese both render well.

Upload images as the Image and the Additional Images, and mention saved Characters and References (outfits, locations, poses, objects) by their @handle. The model sees all of them as one numbered list: your uploads first, then Characters, then References, each in the order you first mention it.

Before the model reads the prompt, Mage turns each mention into its number: @image2 becomes "image 2", @nina becomes "character from image 3", and @trench-coat becomes "outfit from image 4". So you can also write "image 2" yourself, and the model's own label, <Picture 2>, means the same image.

Give every image one role in words, and say what to take from it: the face and hair from one, the coat from another, the room from a third.

With images uploaded, the prompt is an edit, and it works best in three parts, in order. First, what stays: name the image and list everything that must not change, as specifically as you can, such as faces, clothing, objects, the layout, the light, and any text, quoted. Second, what changes, said plainly. Third, the finished image described as a whole, starting with "The target image is…".

The last part matters more than it looks. Describing the result tells the model how the change sits in the picture: where the light falls on it, what it casts or reflects, and what it hides.

When you remove something, say what fills the space: "the bookshelf is gone, and the plain white wall behind it continues to the corner". When you change part of something, say what remains: "his face, sweater, and watch stay; only his pose changes".

MiniMax H3 Image can build one picture from many images at once: an empty room and a set of furniture photos composed into a living room, or a cast of Characters placed in one scene. The more images you add, the more it helps to name each one before you describe the scene.

The model's own long form does this with labels. Define each thing you take ("The green velvet sofa in image 2 is <Subject 2>."), then write one sentence starting with "The target image" that sums up the result, then one short sentence per label saying how much of it to keep, then the new details. Once a label is fully preserved, refer to it only by its label.

An edit can change what kind of picture an image is while keeping what is in it: an untextured 3D render into a photoreal film still, a sketch or a floor plan into a finished room, a game map into a first-person view of the same world. Say what must match (the layout, the objects, their positions), name the new kind of picture, and describe how it looks.

When the viewpoint changes, say where the new camera stands in the source image and which way it looks, then list what it sees from near to far: "seen from the red arrow on the southern lake in image 1, looking north toward the waterfall and the bridge".

Photorealism and fine detail are where MiniMax H3 Image shines, and they come from the specifics a photographer would note: the camera and lens (a 35mm photograph, a 100mm macro, a phone selfie with a wide lens), the light by its source and direction, and the materials, such as worn leather, wet stone, or chipped enamel. Words like "ultra realistic", "8K", or "masterpiece" add nothing a precise description does not.

It paints and draws too. Name the medium exactly, such as a light watercolor, an oil painting with thick brushstrokes, a fine-line ink painting, pixel art, a tarot card, or a stylized 3D render, and describe the subject in that medium.

For a look the base model does not have, add a MiniMax H3 Image LoRA under Model. It starts at a strength of 1; lower it if the style takes over the subject, and put its trigger words in the prompt if its page lists them.

MiniMax H3 Image is tuned to finish in six steps with its guidance built in, so there are no Steps, Guidance Scale, or negative prompt settings. The prompt does all the steering.

Aspect Ratio and Resolution set the size and shape; words such as "16:9" or "4K" in the prompt do nothing useful. Choose the shape first, then describe a composition that fits it.

The Seed fixes the starting noise. With the same Seed and settings, change one sentence at a time to see what each sentence does.