TTangenty/ScaleEdit-labeling-25k

ScaleEdit 25k Fine-Grained Editing Masks

This release contains 25,664 samples that passed both quality and fine-grained editability filtering and reached final mask labeling. It joins source and edited images, instructions, filtering decisions and audits, grounding, and masks. The original ScaleEdit-12M dataset declares CC BY-NC-SA 4.0; this derivative is shared under the same license. A filtered candidate subset was sourced from QingyuShi/scaleedit-filtered-6m.

Contents

• 1,073 nonempty Parquet shards under shards/.
• All original source fields, including source and edited image bytes, instructions, task labels, and provenance.
• quality__, quality_audit__, scene__, scene_audit__, grounding__, and mask__ fields for all persisted stages.
• mask__qc_flag: OK (25,085), MASK_REVIEW (472), or GROUND_FAIL (107). Every row that reached mask labeling is retained; OK is automatic QC, not human verification.
• dataset_manifest.json: full schema, counts, shard index, and source provenance. COMPLETE marks the validated export.

The stable row key is source_shard plus row_idx. Mask labels include mask__mask_png, per-instance mask__instance_masks RLE and audit data, and QC fields.

Loading

from datasets import load_dataset

data = load_dataset(
    "TTangenty/ScaleEdit-labeling-25k",
    data_files={"train": "shards/*.parquet"},
)

Source and citation

Please follow the original dataset's non-commercial and share-alike terms, and cite ScaleEdit-12M and its paper, ScaleEdit-12M: Scaling Open-Source Image Editing Data Generation via Multi-Agent Framework (Chen et al., 2026; arXiv:2603.20644).

@article{chen2026scaleedit,
  title={ScaleEdit-12M: Scaling Open-Source Image Editing Data Generation via Multi-Agent Framework},
  author={Chen, Guanzhou and others},
  journal={arXiv preprint arXiv:2603.20644},
  year={2026}
}