ScaleEdit-12M: Scaling Open-Source Image Editing Data Generation via Multi-Agent Framework

📌 Overview

The largest open-source instruction-based image editing dataset to date.

ScaleEdit-12M contains 12.4 million rigorously verified instruction–image pairs spanning 23 task families across diverse real and synthetic visual domains. It was constructed using ScaleEditor, a fully open-source hierarchical multi-agent framework that eliminates the need for costly proprietary APIs.

🔥 News

• [2026/04/03] 🚀ScaleEdit-12M is released on [Huggingface].
• [2026/03/24] 🔥ScaleEdit-12M paper is released on [arXiv].
• [2026/03/06] 🔥InternVL-U technical report released. Check it out on [arXiv].

✅ TODO

• Release ScaleEdit-12M dataset
• Release ScaleEdit-1M subset
• Release ScaleEditor framework

📊 Dataset Structure

Repository Layout

The dataset is organized into 23 task-specific subdirectories, each containing multiple sharded Parquet files. The directory naming follows the pattern {category_id}_{task_name}:

ScaleEdit-12M/
├── README.md
├── 1.1_style_transfer/                  # Global editing tasks
│   ├── style_transfer_0000.parquet      # (~31.7 GB per shard)
│   ├── style_transfer_0001.parquet
│   ├── ...
│   └── style_transfer_0015.parquet
├── 1.2_tone_adjustment/
│   └── tone_adjustment_XXXX.parquet
├── 1.3_viewpoint_transformation/
├── 1.4_background_replacement/
├── 2.1_object_addition/                 # Object editing tasks
├── 2.2_object_removal/
├── 2.3_object_replacement/
├── 2.4_action_editing/
├── 2.5_part_extraction/
├── 3.1_color_change/                    # Attribute editing tasks
├── 3.2_material_change/
├── 3.3_visual_beautification/
├── 3.4_count_change/
├── 3.5_size_change/
├── 4.1_movie_poster_text_editing/       # Text editing tasks
├── 4.2_gui_interface_text_editing/
├── 4.3_object_surface_text_editing/
├── 4.4_building_surface_text_editing/
├── 5.1_perceptual_reasoning/            # Knowledge-infused tasks
├── 5.2_symbolic_reasoning/
├── 5.3_social_reasoning/
├── 5.4_scientific_reasoning/
└── 6.1_compositional_editing/           # Compositional tasks

Each task folder contains multiple Parquet shards (typically ~31–32 GB each) named {task_name}_{shard_index:04d}.parquet. The number of shards varies by task depending on the volume of data in that category.

Parquet Schema

Each Parquet file contains the following columns:

Example Row

{
    "id": 0,
    "edit_task": "object_addition",
    "edit_instruction": "Add a red and white striped safety barrier at the edge of the platform on the right side of the image.",
    "source_image": <binary bytes>,
    "edited_image": <binary bytes>,
    "source_image_width": 2000,
    "source_image_height": 1500,
    "edited_image_width": 2000,
    "edited_image_height": 1500,
    "instruction_following_score": 3,
    "editing_consistency_score": 3,
    "generation_quality_score": 3
}

The source_image and edited_image columns store images as raw binary bytes. They can be decoded into PIL images:

from PIL import Image
import io

img = Image.open(io.BytesIO(row["source_image"]))

Quality Scores

Every sample has been scored through ScaleEditor's task-aware quality verification mechanism across three dimensions, each rated on a 1–3 scale:

• Instruction Following (IF, 1–3): Does the edited image accurately reflect the intent of the instruction?
• Editing Consistency (EC, 1–3): Are unedited regions preserved? Is the edit spatially coherent with the source?
• Generation Quality (GQ, 1–3): Is the output image free of artifacts, distortions, and visual defects?

In ScaleEdit, only samples with IF=3, EC≥2, GQ≥2 are retained.

🛠️ Highlights

ScaleEdit-12M was constructed using the ScaleEditor framework, which consists of three stages:

1. Source Image Expansion — Curates and expands source images from diverse real and synthetic domains, infusing world knowledge to enable knowledge-grounded editing tasks.
2. Adaptive Multi-Agent Editing — An ensemble of specialized agents generates editing instructions and corresponding edited images, adapting strategies per task family.
3. Task-Aware Quality Verification — A multi-dimensional scoring system evaluates instruction following, editing consistency, and generation quality, filtering out low-quality samples.

Fine-tuning leading foundation models on ScaleEdit-12M yields consistent improvements:

• Up to +10.4% on ImgEdit and +35.1% on GEdit for general editing benchmarks
• Up to +150.0% on RISE and +26.5% on KRIS-Bench for knowledge-infused editing benchmarks

These gains were demonstrated on both UniWorld-V1 and Bagel, showing that open-source agentic pipelines can approach commercial-grade data quality.

🌟 Citation

@article{chen2026scaleedit,
  title={ScaleEdit-12M: Scaling Open-Source Image Editing Data Generation via Multi-Agent Framework},
  author={Chen, Guanzhou and Cui, Erfei and Tian, Changyao and Yang, Danni and Yang, Ganlin and Qiao, Yu and Li, Hongsheng and Luo, Gen and Zhang, Hongjie},
  journal={arXiv preprint arXiv:2603.20644},
  year={2026}
}
@article{tian2026internvl,
  title={InternVL-U: Democratizing Unified Multimodal Models for Understanding, Reasoning, Generation and Editing},
  author={Tian, Changyao and Yang, Danni and Chen, Guanzhou and Cui, Erfei and Wang, Zhaokai and Duan, Yuchen and Yin, Penghao and Chen, Sitao and Yang, Ganlin and Liu, Mingxin and others},
  journal={arXiv preprint arXiv:2603.09877},
  year={2026}
}