---
license: apache-2.0
task_categories:
- image-segmentation
tags:
- aigc-detection
- diffusion-editing
- image-forgery-detection
- diffusion-models
---

# 🖼️ DiffSeg30k -- A multi-turn diffusion-editing dataset for localized AIGC detection

A dataset for **segmenting diffusion-based edits** — ideal for training and evaluating models that localize edited regions and identify the underlying diffusion model, as presented in the paper [DiffSeg30k: A Multi-Turn Diffusion Editing Benchmark for Localized AIGC Detection](https://huggingface.co/papers/2511.19111).

## 📁 Dataset Usage
- `xxxxxxxx.image.png`: Edited images. Each image may have undergone 1, 2, or 3 editing operations.
- `xxxxxxxx.mask.png`: The corresponding mask indicating edited regions, where pixel values encode both the type of edit and the diffusion model used.

Load images and masks as follows:

```python
from datasets import load_dataset
dataset = load_dataset("Chaos2629/Diffseg30k", split="train")
image, mask = dataset[0]['image'], dataset[0]['mask']
```

## 🧠 Mask Annotation

Each mask is a grayscale image (PNG format), where pixel values correspond to a specific editing model. The mapping is as follows:

| Mask Value | Editing Model                                        |
|------------|------------------------------------------------------|
| 0          | background                                           |
| 1          | stabilityai/stable-diffusion-2-inpainting            |
| 2          | kolors                                               |
| 3          | stabilityai/stable-diffusion-3.5-medium              |
| 4          | flux                                                 |
| 5          | diffusers/stable-diffusion-xl-1.0-inpainting-0.1     |
| 6          | glide                                                |
| 7          | Tencent-Hunyuan/HunyuanDiT-Diffusers                 |
| 8          | kandinsky-community/kandinsky-2-2-decoder-inpaint    |

## 📌 Notes

- Each edited image may be edited **multiple turns**, so the corresponding mask may contain several different **label values** ranging from 0 to 8.