ScaleEdit-12M Image Editing Dataset

## Overview ScaleEdit-12M is the largest open-source instruction-based image editing dataset to date, containing approximately 12.4 million rigorously verified instruction–image pairs. The data spans 23 task families across diverse real and synthetic visual domains. It is distributed as Parquet files (tabular + text + image references) and was released in April 2026. ## Schema - instruction — string — natural-language editing instruction - source_image — image/bytes or URI — input image to be edited - target_image — image/bytes or URI — resulting edited image - task_family — string — one of 23 editing task categories (e.g., object removal, style transfer, attribute change) - domain — string — real vs synthetic visual domain label - verification_status — string/bool — quality verification flag from the multi-agent pipeline - +additional metadata columns (agent provenance, generation parameters) ## Sources - HuggingFace: https://huggingface.co/datasets/InternVL-U/ScaleEdit-12M — License: MIT - Associated papers: arXiv:2603.20644, arXiv:2603.09877 ## Methodology The dataset was constructed using ScaleEditor, a fully open-source hierarchical multi-agent framework that eliminates the need for costly proprietary models in data generation. The pipeline generates candidate instruction–image edit triples and applies multi-stage verification agents to filter for quality, instruction faithfulness, and visual coherence before inclusion. Coverage spans 23 task families intended to give broad task diversity across real and synthetic image domains. ## Known gaps & limitations As a synthetically generated dataset, ScaleEdit-12M may inherit biases from the underlying generative and verification models used in the ScaleEditor pipeline. Verification is automated rather than human-graded at scale, so residual noise is expected. Language is English-only for instructions. Domain balance across the 23 task families is not explicitly documented — buyers should validate distribution empirically before using for fine-tuning or eval. Source does not document additional gaps; buyers should validate empirically. ## Intended use & out-of-scope - IS for: training and fine-tuning instruction-based image editing models, building image-editing RAG/retrieval systems, benchmarking edit-instruction following, multimodal model pretraining. - NOT for: evaluation against benchmarks without leakage checks (large synthetic corpus may overlap with common eval suites); safety-critical or medical imaging use without further curation; non-English instruction tasks. _Federated dataset: 261 parquet shards, 6270.10 GB total. Queries and downloads stream through the DataBazaar API._ Original supplier listing: ScaleEdit-12M: Large-Scale Instruction-Based Image Editing Dataset 12.4M verified instruction–image pairs across 23 task families for instruction-based image editing. Largest open-source dataset of its kind, built via the ScaleEditor multi-agent framework. MIT licensed.