Real In, Real Out: What If We Only Use Real Data for Scene Text Editing?
TextRIRO learns to edit scene text by turning real images into their own training pairs.
The method combines character-column shuffling with cropping to disrupt text content while retaining style cues, then trains a conditional diffusion model on those pairs. The authors report strong accuracy and visual realism on real-world benchmarks, with fewer problems from the gap between synthetic training data and real images. They also built TextRIRO-3M for long-text recognition by concatenating edited and original images. Training recognition models with that dataset significantly improved accuracy, according to the paper.
HF Daily Papers' note
The method combines character-column shuffling with cropping to disrupt text content while retaining style cues, then trains a conditional diffusion model on those pairs. The authors report strong accuracy and visual realism on real-world benchmarks, with fewer problems from the gap between synthetic training data and real images. They also built TextRIRO-3M for long-text recognition by concatenating edited and original images. Training recognition models with that dataset significantly improved accuracy, according to the paper.
HF Daily Papers' note
score 4