live
MODEL IDminimax:h3@image

MiniMax H3 Image

MiniMax
by

MiniMax H3 Image is an open-weights image generation and editing model from MiniMax. It creates images from text or edits and combines up to 16 reference images, at 1K, 2K or 4K, or at a custom width and height up to 20 megapixels. It follows counts, positions and layouts closely and renders long, dense text, which suits product and marketing visuals, posters and infographics, concept art, and compositions that merge several subjects.

MiniMax H3 Image

Prompting

How to write text-to-image prompts for MiniMax H3 Image: sizing by preset, exact dimensions or no size, and planning the task and layout before the details.

Introduction

MiniMax H3 Image generates an image from a prompt and an optional size, and it follows instructions closely: how many subjects, where each one stands, which way it faces and which part of the frame stays empty. Write the prompt as a plan of the image, starting with what it is for, then the layout, then the details and the light.

Every count and position in the prompt is an instruction. It names which two desks stay empty, the left pair in the back row, and those are the two the render leaves free.

This guide covers the request, the three ways to size it, and how to structure a prompt so each decision in the image is yours. Quoted text, infographics and UI screens have their own guide, text and layouts. Prompt optimizer covers turning off the prompt rewrite, and editing images and multi-reference composition start from images you already have.

The request

A text-to-image request needs only the prompt, and the size is optional. This is the request behind the hero:

Try in Playground
import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'minimax:h3@image',
  positivePrompt: 'A wide hero illustration for a project management app\'s landing page. An isometric 3D diorama of a small open-plan office built on a single square floor slab, floating on a plain pale lavender background and seen from the front-right corner. The slab sits in the right half of the frame, and the left half is empty pale lavender with nothing in it, kept clear for a headline. On the slab, six white desks stand in two rows of three, each with one open laptop and one coral desk chair. Four people work at the desks, and the two desks on the left of the back row are empty, their chairs pushed in. A tall potted monstera stands in the back-left corner of the slab, a low white bookshelf runs along the back-right edge, and a round mustard rug lies under the front row. Smooth matte clay materials in white, lavender, coral and mustard, soft light from the upper left, gentle ambient occlusion where the desks meet the floor. Clean 3D product illustration, no text, no logos.',
  width: 2560,
  height: 1440
})
import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "minimax:h3@image",
            "positivePrompt": "A wide hero illustration for a project management app's landing page. An isometric 3D diorama of a small open-plan office built on a single square floor slab, floating on a plain pale lavender background and seen from the front-right corner. The slab sits in the right half of the frame, and the left half is empty pale lavender with nothing in it, kept clear for a headline. On the slab, six white desks stand in two rows of three, each with one open laptop and one coral desk chair. Four people work at the desks, and the two desks on the left of the back row are empty, their chairs pushed in. A tall potted monstera stands in the back-left corner of the slab, a low white bookshelf runs along the back-right edge, and a round mustard rug lies under the front row. Smooth matte clay materials in white, lavender, coral and mustard, soft light from the upper left, gentle ambient occlusion where the desks meet the floor. Clean 3D product illustration, no text, no logos.",
            "width": 2560,
            "height": 1440
        })


asyncio.run(main())
curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "imageInference",
      "taskUUID": "7e2c4a91-5b3f-4d8e-a1c6-9f0b3d5e7a24",
      "model": "minimax:h3@image",
      "positivePrompt": "A wide hero illustration for a project management app's landing page. An isometric 3D diorama of a small open-plan office built on a single square floor slab, floating on a plain pale lavender background and seen from the front-right corner. The slab sits in the right half of the frame, and the left half is empty pale lavender with nothing in it, kept clear for a headline. On the slab, six white desks stand in two rows of three, each with one open laptop and one coral desk chair. Four people work at the desks, and the two desks on the left of the back row are empty, their chairs pushed in. A tall potted monstera stands in the back-left corner of the slab, a low white bookshelf runs along the back-right edge, and a round mustard rug lies under the front row. Smooth matte clay materials in white, lavender, coral and mustard, soft light from the upper left, gentle ambient occlusion where the desks meet the floor. Clean 3D product illustration, no text, no logos.",
      "width": 2560,
      "height": 1440
    }
  ]'
runware run minimax:h3@image \
  positivePrompt="A wide hero illustration for a project management app's landing page. An isometric 3D diorama of a small open-plan office built on a single square floor slab, floating on a plain pale lavender background and seen from the front-right corner. The slab sits in the right half of the frame, and the left half is empty pale lavender with nothing in it, kept clear for a headline. On the slab, six white desks stand in two rows of three, each with one open laptop and one coral desk chair. Four people work at the desks, and the two desks on the left of the back row are empty, their chairs pushed in. A tall potted monstera stands in the back-left corner of the slab, a low white bookshelf runs along the back-right edge, and a round mustard rug lies under the front row. Smooth matte clay materials in white, lavender, coral and mustard, soft light from the upper left, gentle ambient occlusion where the desks meet the floor. Clean 3D product illustration, no text, no logos." \
  width=2560 \
  height=1440
{
  "taskType": "imageInference",
  "taskUUID": "7e2c4a91-5b3f-4d8e-a1c6-9f0b3d5e7a24",
  "model": "minimax:h3@image",
  "positivePrompt": "A wide hero illustration for a project management app's landing page. An isometric 3D diorama of a small open-plan office built on a single square floor slab, floating on a plain pale lavender background and seen from the front-right corner. The slab sits in the right half of the frame, and the left half is empty pale lavender with nothing in it, kept clear for a headline. On the slab, six white desks stand in two rows of three, each with one open laptop and one coral desk chair. Four people work at the desks, and the two desks on the left of the back row are empty, their chairs pushed in. A tall potted monstera stands in the back-left corner of the slab, a low white bookshelf runs along the back-right edge, and a round mustard rug lies under the front row. Smooth matte clay materials in white, lavender, coral and mustard, soft light from the upper left, gentle ambient occlusion where the desks meet the floor. Clean 3D product illustration, no text, no logos.",
  "width": 2560,
  "height": 1440
}
Response
[
  {
    "taskType": "imageInference",
    "taskUUID": "7e2c4a91-5b3f-4d8e-a1c6-9f0b3d5e7a24",
    "imageUUID": "c41f8e2d-6a9b-4c3e-8d07-2b5a9e1f4c63",
    "imageURL": "https://im.runware.ai/image/os/a14d18/ws/2/ii/c41f8e2d-6a9b-4c3e-8d07-2b5a9e1f4c63.jpg"
  }
]
  • positivePrompt takes 1 to 32,000 characters. With settings.promptOptimizer on, which is the default, MiniMax rewrites the prompt into a detailed scene description before generating. The response does not include the rewritten version.
  • The size comes from resolution or from width and height, never both.
  • width and height go together, and a request with only one of them is rejected.

MiniMax H3 Image Highspeed takes the same parameters and limits, so any request in this guide runs on it with only model changed to minimax:h3@image-highspeed. It generates faster and bills its image tokens at a higher rate.

Choosing the size

The three ways to size a request answer different questions. A preset sets how many pixels and leaves the shape to the model, width and height set exactly which frame, and leaving the size out returns 4K.

Resolution presets

resolution takes 1K, 2K or 4K, and the model sets the width and height itself. With no reference image it picks the aspect ratio from the prompt, so a preset request should name its orientation in words. The fitness story below asked for "a vertical 9:16 story image" at 2K and sent no dimensions.

Read the dimensions from the file. This 9:16 request came back at 1536 × 2720, close to 9:16 but not exact, so a layout that needs the exact ratio should send width and height. With reference images, the preset follows their ratio instead, which editing images covers.

Exact sizes

Send width and height when the image fills a fixed slot, such as a link preview or a banner. The pair is generated exactly as given, never scaled or rounded, so a 1200 × 630 request returns a 1200 × 630 file with no crop step. It has to fit three limits:

  • The area is between 655,360 and 20,000,000 pixels.
  • Each side is between 468 and 7,745 pixels.
  • The aspect ratio is between 1:3 and 3:1.

The area floor rules out small renders. It equals 1024 × 640, so 1080 × 1080 passes and 800 × 800 is rejected. At the top, the largest square is 4472 × 4472.

Choose the size before writing the layout. Positions such as "the right third" and "the left half" are measured against the frame you send, so a layout written for this wide card has to be rewritten for a vertical slot.

Leaving the size out

A request with no resolution, width or height comes back at 4K, at a ratio the model picks. Name the orientation in the prompt the same way as with a preset.

4K uses about four times the image tokens of 1K or 2K, and those two use about the same. Send 2K when the image will never be shown larger, and leave the size out when you want the full 4K.

Structuring the prompt

MiniMax's advice for this model is to put the task before the style and the layout before the details. The two product tiles below share a subject and a size. The first prompt is one line, and the second plans the image in that order.

Both went through the prompt optimizer, which expands a short prompt into a full scene before generating, so the one-line version still returns a finished product shot. The shoe colors, the way they face, the camera height and the background in it are choices MiniMax made for you.

The structured prompt, clause by clause:

A square product tile for an online running store's new arrivals page, Three running shoes stand in a single row on a floating pale concrete shelf, evenly spaced, shown in side profile with every toe pointing left, from left to right in coral, slate blue and off-white, The camera is at shelf height, looking straight on, The row sits in the lower half of the frame, and the upper half is a plain warm gray wall with nothing on it, Each shoe has an engineered knit upper with a visible weave, a thick foam midsole with a matte finish and a rubber outsole with a fine tread, Soft diffused light from above and slightly left, a narrow highlight along each midsole, short soft contact shadows on the shelf. Photoreal e-commerce product photography, no text, no logos
TaskLayoutViewing angleEmpty spaceDetailsStyleExclusions

Name the task first. "A square product tile for an online running store's new arrivals page" says what the image is and where it will run before any detail does. A catalog tile and a lifestyle ad of the same shoes call for different framing, and the opening clause is where that gets decided.

Set the layout before the details. The count, the order of the colors, the direction the toes point, the camera height and the empty upper half are all settled before a word about knit or foam. Dense layouts such as an infographic follow the same order section by section, which text and layouts covers.

Describe the style through what is visible. "Soft diffused light from above and slightly left, a narrow highlight along each midsole" names a light source and the surface it falls on, and the model can render both. A word like "premium" leaves the model to guess what premium looks like.

Keep the requirements free of contradictions. "Minimalist" next to a list of ten props, or "overcast" next to "hard midday shadows", asks for two images in one prompt. Lead with the details that must be right and cut the ones that do not matter.

Counts and positions

Write every count as a number and every position against the frame. The team photo below lists five people from left to right, one clause each, so every outfit and prop belongs to one person.

Anchor relationships to the frame as well. The hand on the shoulder is aimed at "the man in the middle", a place in the frame. "The man to her left" would depend on which way she faces.

Tips

  1. Open with the task. Name what the image is and where it will run before any detail.

  2. Plan the layout before the details. Count, placement, viewing angle and empty space come first, then materials and light.

  3. Write counts as numbers and positions against the frame. "Two rows of three" and "second from the left" leave nothing for the model to settle, and "her left" depends on which way she faces.

  4. Describe style in visible terms. Name the light and the materials instead of "premium" or "high-end".

  5. Cut contradictions. Two clauses that disagree ask for two images. Keep the one that matters.

  6. End with what to leave out. Exclusions such as "no text, no logos" go in the prompt itself, as its last clause.

  7. Name the orientation when the model picks the shape. With a preset or no size at all, the prompt decides the aspect ratio.

  8. Send exact dimensions for fixed slots. width and height come back exactly as sent, so the file needs no crop.

  9. Send a preset below 4K when you can. 1K and 2K use about the same image tokens, and a request with no size returns 4K at about four times as many.