SIGMA: Full 1.1M Image Manipulation Localization Dataset

This repository is the canonical full release of the SIGMA-annotated dataset. It contains 1,063,140 samples (approximately 1.1 million) across 24 editing-operation categories, converted from four public text-driven image-editing corpora:

• AnyEdit
• CrispEdit-2M
• OmniEdit-Filtered-1.2M
• PromptfixData

Two smaller convenience subsets of the same annotations are available on ModelScope:

• 10,000 samples: PeiyuZhuang/SIGMA_small
• 100,000 samples: PeiyuZhuang/SIGMA_partial

The 10K and 100K releases are subsets of this repository, not separate datasets. Use this full release for large-scale training.

Dataset Summary

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 image manipulation localization (IML) models. SIGMA (Semantic-Difference Instruction-Grounding Mask Annotator) converts these existing records into IML supervision.

SIGMA performs semantic-feature differencing between the original and edited images and injects an instruction-derived spatial prior through bidirectional cross-modal refinement. This combination helps suppress diffusion-induced perturbations while retaining both intended edits and editor side-effects that instruction-only grounding may miss.

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.

Validation and Downstream Utility

Across CoCoGlide, AutoSplice, MagicBrush, DEAL-300K, and OpenSDI, SIGMA reaches 89.83% average F1 and 83.52% average IoU for automatic mask generation. This is an improvement of 12.20 F1 points and 11.16 IoU points over DDPM-CD. On DEAL-300K, SIGMA reaches 76.43% F1, compared with 20.90% for DDPM-CD.

The annotations are also useful as training supervision. Fine-tuning CAT-Net, MVSS-Net, PSCC-Net, and TruFor on SIGMA-annotated editing data improves the average F1 of all four detectors over their public weights. The four gains range from 4.58 to 30.79 points, with a mean gain of 20.98 points across five held-out datasets.

Dataset Family

• Full release (this repository): 1,063,140 samples from four source corpora and 24 operation categories.
• 100K subset: selected from this full release for medium-scale experiments.
• 10K subset: 2,500 samples from each source corpus, intended for quick inspection and pipeline testing.

No additional annotation method is used for the subsets.

Repository Layout

SIGMA.zip.001
SIGMA.zip.002
...
SIGMA.zip.113
SIGMA_instructions/
  AnyEdit.jsonl
  CrispEdit-2M.jsonl
  OmniEdit-Filtered-1.2M.jsonl
  PromptfixData.jsonl

SIGMA.zip.001 through SIGMA.zip.113 are sequential fragments of the full ZIP archive. They must be downloaded completely and concatenated in numeric order before extraction. The JSONL files under SIGMA_instructions/ provide the editing-instruction manifests grouped by source corpus.

Download and Extraction

Download with the Hugging Face CLI:

hf download PatrickZZZ/SIGMA \
  --repo-type dataset \
  --local-dir SIGMA
cd SIGMA

Alternatively, use Git LFS:

git lfs install
git clone https://huggingface.co/datasets/PatrickZZZ/SIGMA
cd SIGMA

Reassemble the numbered fragments with Python:

from pathlib import Path

parts = sorted(Path(".").glob("SIGMA.zip.[0-9][0-9][0-9]"))
if len(parts) != 113:
    raise RuntimeError(f"Expected 113 archive fragments, found {len(parts)}")

with open("SIGMA.zip", "wb") as output:
    for part in parts:
        print(f"Appending {part}")
        with open(part, "rb") as source:
            while chunk := source.read(1024 * 1024):
                output.write(chunk)

Then extract the reconstructed archive:

unzip SIGMA.zip

On Windows, the same Python reassembly script can be used before running:

Expand-Archive SIGMA.zip

Intended Uses

Appropriate uses include:

• training IML models on text-driven edits;
• studying generalization across image editors and generators;
• investigating learning from automatically generated pixel masks;
• comparing annotation and data-curation strategies for image forensics.

The ModelScope subsets are recommended for pipeline debugging before downloading this full release.

Limitations

• The masks are generated automatically and can contain residual errors.
• Ambiguous or compositional instructions may be parsed incompletely.
• Extremely subtle edits and unintended editor side-effects remain difficult.
• The source corpora and operation distribution do not cover every image editor, manipulation type, or deployment domain.
• This dataset should not replace manually annotated benchmarks when reporting final mask-quality results.

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
• CrispEdit-2M: CC BY 4.0
• OmniEdit-Filtered-1.2M: MIT

Source repositories:

• AnyEdit
• CrispEdit-2M
• OmniEdit-Filtered-1.2M
• PromptfixData

Citation

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

@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}
}

The source-corpus papers are:

• AnyEdit: AnyEdit: Mastering Unified High-Quality Image Editing for Any Idea
• CrispEdit-2M: EditMGT: Unleashing Potentials of Masked Generative Transformers in Image Editing
• OmniEdit-Filtered-1.2M: OmniEdit: Building Image Editing Generalist Models Through Specialist Supervision
• PromptfixData: PromptFix: You Prompt and We Fix the Photo

Related Resources

• Project page
• Paper
• Code
• Model checkpoint