---
title: MiniMax-H3-Image-Self-Generated-Dataset
canonical_url: "https://www.modelscope.cn/datasets/DiffSynth-Studio/MiniMax-H3-Image-Self-Generated-Dataset"
md_url: "https://www.modelscope.cn/datasets/DiffSynth-Studio/MiniMax-H3-Image-Self-Generated-Dataset.md"
repository: DiffSynth-Studio/MiniMax-H3-Image-Self-Generated-Dataset
last_updated: 2026-11-13
license: apache-2.0
storage_size: "152 GB"
downloads: 80670
stars: 0
---

# MiniMax-H3-Image-Self-Generated-Dataset

> MiniMax-H3-Image-Self-Generated-Dataset - DiffSynth-Studio 在 ModelScope 开源的数据集。MiniMax-H3-Image Self-Generated Dataset

DiffSynth-Studio/MiniMax-H3-Image-Self-Generated-Dataset 是 ModelScope 魔搭社区上的数据集，存储大小 152 GB，采用 apache-2.0 许可。

- **Repository**: DiffSynth-Studio/MiniMax-H3-Image-Self-Generated-Dataset
- **License**: apache-2.0
- **Storage size**: 152 GB
- **Downloads**: 80670
- **Stars**: 0
- **Last updated**: 2026-11-13

Source: https://www.modelscope.cn/datasets/DiffSynth-Studio/MiniMax-H3-Image-Self-Generated-Dataset

---

# MiniMax-H3-Image Self-Generated Dataset

A synthetic text-to-image dataset of **39,525 images** produced with
[MiniMax-H3-Image](https://modelscope.cn/models/MiniMax/MiniMax-H3-Image) through
[DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio).

| Subset | Images | Task | Metadata |
|---|---|---|---|
| **T2I** | 39,525 | text → image | `metadata.jsonl` |

## Structure

```
.
├── metadata.jsonl     # all samples
├── images/            # generated images, flat naming
│   ├── 000001.jpg
│   └── ... 039525.jpg
└── stats.json         # resolution distribution and aggregate statistics
```

## Metadata

Each line of `metadata.jsonl` describes one image:

```json
{
  "image": "images/000001.jpg",
  "prompt": "A cyberpunk netrunner is centered in the frame in a sharp medium shot, facing the camera with focused eyes, ..."
}
```

| Field | Type | Description |
|---|---|---|
| `image` | str | Path to the generated image, relative to the dataset root |
| `prompt` | str | The expanded caption that was fed to the model — not the original short prompt |

## Generation settings

| Setting | Value |
|---|---|
| Model | MiniMax-H3-Image, pruned BF16 DiT + EntroPack pre-quantized text encoder |
| Inference steps | 6 |
| CFG scale | 1.0 |
| Area budget | 4096 × 4096 = 16,777,216 pixels |
| Dimension rule | keep the chosen aspect ratio, round each side down to a multiple of 32 |
| Aspect ratios | one of 1:1, 5:4, 4:3, 3:2, 16:9, 21:9, 4:5, 3:4, 2:3, 9:16 |
| Seed | per-sample |
| VAE decode | tiled |

### Resolutions

Each image keeps the aspect ratio chosen during prompt expansion and is fitted into a
`4096 × 4096` area budget, with both dimensions rounded down to a multiple of 32. Average
area is 16.63 megapixels.

| Resolution (W×H) | Aspect | Images |
|---|---|---|
| 4992 × 3328 | 3:2 | 11,811 |
| 3520 × 4704 | 3:4 | 8,536 |
| 5440 × 3072 | 16:9 | 4,826 |
| 3328 × 4992 | 2:3 | 4,597 |
| 4096 × 4096 | 1:1 | 4,100 |
| 4704 × 3520 | 4:3 | 2,886 |
| 3648 × 4576 | 4:5 | 2,410 |
| 3072 × 5440 | 9:16 | 247 |
| 6240 × 2656 | 21:9 | 94 |
| 4576 × 3648 | 5:4 | 18 |
