---
title: SIGMA_small
canonical_url: "https://www.modelscope.ai/datasets/PeiyuZhuang/SIGMA_small"
md_url: "https://www.modelscope.ai/datasets/PeiyuZhuang/SIGMA_small.md"
repository: PeiyuZhuang/SIGMA_small
last_updated: 2026-10-19
license: other
storage_size: "10.0 GB"
downloads: 2
stars: 0
---

# SIGMA_small

> SIGMA_small - An open-source dataset by PeiyuZhuang on ModelScope. SIGMAsmall: 10K Subset of the SIGMA-Annotated IML Dataset

PeiyuZhuang/SIGMA_small is a dataset on ModelScope. totalling 10.0 GB. licensed under other.

- **Repository**: PeiyuZhuang/SIGMA_small
- **License**: other
- **Storage size**: 10.0 GB
- **Downloads**: 2
- **Stars**: 0
- **Last updated**: 2026-10-19

Source: https://www.modelscope.ai/datasets/PeiyuZhuang/SIGMA_small

---

# SIGMA_small: 10K Subset of the SIGMA-Annotated IML Dataset

**SIGMA_small contains 10,000 samples selected from the full
[SIGMA dataset](https://huggingface.co/datasets/PatrickZZZ/SIGMA).**
It is a compact convenience subset for quickly inspecting the data, testing
input pipelines, and running small-scale image manipulation localization (IML)
experiments. It is not a separate dataset and introduces no annotations beyond
the full SIGMA release.

The subset contains **2,500 samples from each of the four source corpora**:
AnyEdit, CrispEdit-2M, OmniEdit-Filtered-1.2M, and PromptfixData.

## Dataset Family

- **This repository:** 10,000-sample subset  
  [ModelScope: PeiyuZhuang/SIGMA_small](https://www.modelscope.ai/datasets/PeiyuZhuang/SIGMA_small)
- **Larger subset:** 100,000 samples  
  [ModelScope: PeiyuZhuang/SIGMA_partial](https://www.modelscope.ai/datasets/PeiyuZhuang/SIGMA_partial)
- **Canonical full release:** approximately 1.1 million samples  
  [Hugging Face: PatrickZZZ/SIGMA](https://huggingface.co/datasets/PatrickZZZ/SIGMA)

Use the full Hugging Face release for large-scale training and the two
ModelScope subsets when a smaller download is more practical.

## What Is SIGMA?

Text-driven editing corpora already contain millions of
`(original image, edited image, editing instruction)` records, but they do not
provide the pixel-level masks required to train IML models. SIGMA
(*Semantic-Difference Instruction-Grounding Mask Annotator*) converts these
records into IML supervision by combining semantic-feature differencing with an
instruction-derived spatial prior.

Each released sample is derived from a public editing record and adds a binary
localization mask generated by SIGMA. These masks are automatic annotations,
not human-drawn ground truth.

Across five public benchmarks, SIGMA reaches **89.83% average F1** and
**83.52% average IoU** for mask generation. The full SIGMA-annotated release
contains **1,063,140 samples** spanning **24 editing-operation categories**.
Fine-tuning four existing IML detectors on SIGMA annotations raises their
average F1 by **20.98 points on average** across five held-out datasets.

## Files and Integrity

- `SIGMA_small.zip`: all 10,000 samples in this subset
- `SHA256SUMS`: SHA-256 checksum for the archive

After downloading, verify the archive before extraction:

```bash
sha256sum -c SHA256SUMS
unzip SIGMA_small.zip
```

On Windows PowerShell:

```powershell
(Get-FileHash SIGMA_small.zip -Algorithm SHA256).Hash.ToLower()
Get-Content SHA256SUMS
Expand-Archive SIGMA_small.zip
```

## Download

:modelscope-code[]{type="sdk"}

:modelscope-code[]{type="git"}

Because the archive is stored with Git LFS, command-line Git users should
install Git LFS before cloning:

```bash
git lfs install
git clone https://www.modelscope.ai/datasets/PeiyuZhuang/SIGMA_small.git
```

## Intended Uses

Appropriate uses include:

- testing SIGMA data loading and preprocessing code;
- small-scale IML training or debugging;
- studying text-driven image manipulation localization;
- evaluating methods that learn from automatically generated masks.

This subset is not intended to replace manually annotated evaluation
benchmarks. Automatic masks can contain residual errors, especially for
ambiguous instructions, extremely subtle edits, and unintended editor
side-effects.

## License and Terms of Use

**The SIGMA-annotated data are released for academic research only.** This
repository does not replace or override the licenses of the source corpora.
Users must comply with the applicable source-dataset terms and cite both SIGMA
and the corresponding source dataset.

- **PromptfixData:** research use only; the dataset and models trained on it may
  not be used commercially. Users must also follow the licenses of its source
  data.
- **AnyEdit:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
- **CrispEdit-2M:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
- **OmniEdit-Filtered-1.2M:** [MIT](https://opensource.org/license/mit/)

Source repositories:

- [AnyEdit](https://huggingface.co/datasets/Bin1117/AnyEdit)
- [CrispEdit-2M](https://huggingface.co/datasets/WeiChow/CrispEdit-2M)
- [OmniEdit-Filtered-1.2M](https://huggingface.co/datasets/TIGER-Lab/OmniEdit-Filtered-1.2M)
- [PromptfixData](https://huggingface.co/datasets/yeates/PromptfixData)

## Citation

Please cite SIGMA and the source corpus associated with each sample:

```bibtex
@misc{zhuang2026sigma,
  title         = {SIGMA: Semantic-Difference Instruction-Grounding Mask Annotator for Text-Driven Image Manipulation Localization},
  author        = {Zhuang, Peiyu and Yang, Jianquan and Li, Haodong and Cai, Zhuoying and Xie, Ruitao and Zeng, Jishen and Chen, Baoying and Huang, Jiwu and Cao, Xiaochun},
  year          = {2026},
  eprint        = {2605.27924},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2605.27924}
}
```

## Related Resources

- [Project page](https://zhuangpeiyu.github.io/projects/SIGMA/)
- [Paper](https://arxiv.org/abs/2605.27924)
- [Code](https://github.com/ZhuangPeiyu/SIGMA)
- [Model checkpoint](https://huggingface.co/PatrickZZZ/SIGMA)