EditNet
Introducing EditNet — a high-quality before/after dataset we built to train our models, named in homage to ImageNet.
Bootstrapped with real life photoshoots
To kickstart our dataset, we went out and shot thousands of before/after pairs in the wild. The result: a diverse collection covering hundreds of object categories, across different locations and lighting conditions.
Natural images only
We curated millions of photos chosen for their aesthetic quality, all shot with professional lighting and composition. Synthetic and 3D imagery were used only to fill gaps where real images were unavailable.
Enhanced with AI
We trained a suite of high-quality specialist models — such as high-resolution background removal and object erasing — on our manually collected dataset. These models were then used to process the curated data.
Fully annotated
We collected comprehensive annotations — including categories, bounding boxes, masks, viewpoints, occlusions, and perspective — enabling us to generate diverse training data variants and measure impact through ablation studies.
Human Reviewed
Every annotation and AI-processed image has been carefully validated by trained human experts to ensure accuracy and consistency.
