Teaching AI to edit text in photos without breaking everything else

Scene text editing—changing words in photos while keeping images realistic—is hard because models must preserve backgrounds and non-target areas. TextSculptor provides a 3.2M-sample dataset built from automated synthesis and real compositing, plus TextSculpt-Bench, a standardized test covering text addition, replacement, removal, and mixed edits. The framework includes OCR-based evaluation to measure both text accuracy and visual quality. Results show open-source models can now match proprietary systems.