TextSculptor: Training and Benchmarking Scene Text Editing

Researchers introduced TextSculptor, a framework that automates scene text editing data generation.

They released TextSculpt-Data, a 3.2 million-sample dataset covering text addition, replacement, removal, and hybrid editing.

They also launched TextSculpt-Bench to evaluate text accuracy, visual quality, and background preservation.

The effort targets two key bottlenecks in the field: limited training data and weak evaluation standards.

The release aims to help open-source models close the gap with proprietary systems in scene text editing.