TextRIRO trains scene text editing on real self-paired images
Original title: Real In, Real Out: What If We Only Use Real Data for Scene Text Editing?
Shuffle-and-crop self-pairs train a diffusion editor, and derived TextRIRO-3M lifts long-text recognition.
Researchers propose TextRIRO, a conditional diffusion baseline for scene text editing that turns each real image into its own training pair. Character-column shuffling plus cropping removes text meaning while keeping style cues, avoiding synthetic paired data. They also build TextRIRO-3M by concatenating edited and original images and report higher recognition accuracy. Code is on GitHub.
Unconfirmed: Accuracy gains are reported by the authors and not independently verified.
- Self-pairs disrupt images with crop, shuffle, and mask.
- Shuffle&Crop is the mix that best keeps style cues.
- TextRIRO-3M is built for long-text recognition training.
- Authors say the method narrows the synthetic-to-real gap.
From the original article
1 min read at huggingface.co
Aiming to modify textual content while preserving the original style, Scene Text Editing (STE) has significant practical value in many applications. STE typically requires paired images before and after editing for training. However, recent approaches rely heavily on synthetic paired data and complex disentanglement strategies.
By Xingsong Ye, Yongkun Du, Jiaxin Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.