# MiniMax H3 Image

Source: https://docs.prodia.com/models/minimax-h3-image/

MiniMax H3 Image is MiniMax's image generation and editing model from the MiniMax H3 family. It generates images of up to 20 megapixels natively, without an upscaler, at any size and aspect ratio from 1:3 to 3:1, and it edits or combines up to 16 reference images in a single request. It follows long, detailed prompts closely and renders text well, which makes it a strong choice for high-resolution photography, posters and layouts, and multi-image composition.

### Architecture

MiniMax H3 Image uses the diffusion transformer of MiniMax H3 (33.1B parameters, a single-stream DiT) without its audio branch. The prompt and the reference images are encoded together by the MiniMax H3 vision-language text encoder, so the model reads references as part of the instruction rather than only as pixels. A dedicated image VAE with a transformer decoder produces the final image. Text-to-image and editing are one model: an edit is the same generation, conditioned on the reference images.

### When to use MiniMax H3 Image

| Scenario                                                                | Job type                                     |
| ----------------------------------------------------------------------- | -------------------------------------------- |
| High-resolution images (4–20 MP) with fine detail                       | `inference.minimax.h3-image.base.txt2img.v0` |
| Exact output sizes, including tall or wide formats up to 3:1            | `inference.minimax.h3-image.base.txt2img.v0` |
| Posters, covers and layouts with readable text                          | `inference.minimax.h3-image.base.txt2img.v0` |
| Editing a photo while keeping its subject and composition               | `inference.minimax.h3-image.base.img2img.v0` |
| Composing a scene from several references (products, characters, rooms) | `inference.minimax.h3-image.base.img2img.v0` |

For sub-second generation at 1–4 MP, consider [FLUX.2 Klein](https://docs.prodia.com/models/flux-2-klein/). For vector output, use [Recraft V4](https://docs.prodia.com/models/recraft-v4/).

### Job types

| Job type                                     | Description                           | ETA                                   |
| -------------------------------------------- | ------------------------------------- | ------------------------------------- |
| `inference.minimax.h3-image.base.txt2img.v0` | Generate an image from a text prompt  | \~2s at 4 MP, \~6s at 16 MP           |
| `inference.minimax.h3-image.base.img2img.v0` | Edit or combine 1–16 reference images | \~2–3s for one image; \~20–75s for 16 |

> note:
>
> Edits with many reference images can take over a minute. Use the [async API](https://docs.prodia.com/reference/async/) for those to avoid holding a connection open.

### Parameters

*Common to all:*

- `prompt` (required) — text description or editing instruction, 1–32,000 characters. The model reads the first 4096 tokens.
- `width`, `height` — exact output size in pixels, set together. Each side 468–7745 px, an area of 655,360–20,000,000 pixels, and an aspect ratio between 1:3 and 3:1.
- `seed` — integer 0–2147483647 for reproducible results.

*Text-to-image (`txt2img`):*

- `width` and `height` default to 2048×2048.

*Image editing (`img2img`):*

- `images` — 1–16 input images, in the order the prompt refers to them. Each side 128–8192 px, with the long side at most 4 times the short side; images above 4 MP are downscaled to 4 MP. If omitted, the uploaded files are used in upload order.
- Without `width` and `height`, the output keeps the first image's aspect ratio (limited to 1:3–3:1) at 2048×2048 worth of pixels.

### Prompting tips

- *Describe the whole image:* the model responds best to long, complete English descriptions covering subject, setting, composition, lighting and style, rather than short requests.
- *Put rendered text in quotes:* for signs, titles and labels, quote the exact text and say where it goes, for example `a poster with the title "BLUE NOTE NIGHTS" at the top`.
- *Go big for detail:* the model is at its best between 4 and 20 MP; around 16 MP (for example 4096×4096 or 5472×3072) gives the most detail.
- *Name the references:* refer to inputs as Image 1, Image 2, and so on (or `<Picture 1>`, `<Picture 2>`), in upload order.
- *Structure edits:* say what to keep, then what to change, then describe the whole target image. Put several changes into one edit rather than chaining edits.

### Examples

A wide, high-resolution photograph:

```json
{
  "type": "inference.minimax.h3-image.base.txt2img.v0",
  "config": {
    "prompt": "A wide photograph of a lighthouse on a rocky coast at sunset, waves crashing against the rocks, warm golden light, seagulls in the sky, long exposure softening the sea",
    "width": 5472,
    "height": 3072,
    "seed": 42
  }
}
```

A poster with rendered text:

```json
{
  "type": "inference.minimax.h3-image.base.txt2img.v0",
  "config": {
    "prompt": "A vintage jazz festival poster on a deep navy background with gold Art Deco sunbursts. The title \"BLUE NOTE NIGHTS\" in large gold serif letters at the top, a saxophone player in silhouette in the center, and the line \"OCTOBER 18-20\" at the bottom",
    "width": 3072,
    "height": 4096
  }
}
```

Combining two reference images (upload `room.jpg` and `lamp.png` with the job):

```json
{
  "type": "inference.minimax.h3-image.base.img2img.v0",
  "config": {
    "prompt": "Keep the room in Image 1 exactly as it is, and place the brass floor lamp from Image 2 in the empty corner on the right, lit and casting a warm glow on the wall",
    "images": ["room.jpg", "lamp.png"]
  }
}
```

### Guides

[Generating Images](https://docs.prodia.com/guides/generating-images/) — Step-by-step guide for generating images with Prodia.

[Transforming Images](https://docs.prodia.com/guides/transforming-images/) — Guide for editing and transforming images with Prodia.

[Pricing](https://docs.prodia.com/reference/pricing/#minimax-h3-image) — Per-image prices by output size and number of input images.