AI Paper Atlas

Research map / Image generation / Text-to-image diffusion

Real In, Real Out: What If We Only Use Real Data for Scene Text Editing?

Xingsong Ye, Yongkun Du, Jiaxin Zhang, Chong Sun, Zhineng Chen

NeurIPS 2026 · Image generation › Text-to-image diffusion · approach: Personalized and region-controlled generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Investigates whether scene text editing can be trained using only real (non-synthetic) paired image data instead of synthetic paired data and complex disentanglement strategies.

Problem

Scene Text Editing (modifying text content in natural images while preserving style) typically needs paired before/after images, which current methods obtain mainly from synthetic data; this synthetic data struggles to capture the diversity and complexity of natural scenes.

Approach

Uses only real paired data for training scene text editing, arguing that real data suffices to avoid the domain gap of synthetic pairs, as suggested by the title; the specific mechanism is not stated in the provided abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Train only on real self-paired data (Shuffle&Crop) instead of synthetic tripletsTextRIRO · Train-on-edits STR proxy avgaccuracy (%): 76.89 -> 86.67Improved
Use Shuffle&Crop instead of Crop only for straight text (ablation)TextRIRO · IC13 equal-lengthaccuracy (%): 77.83 -> 80.16Improved
Use Crop only instead of Shuffle&Crop for curved text (ablation)TextRIRO · CUTE equal-lengthaccuracy (%): 58.42 -> 60.93Mixed
Remove self-pairing perturbation (identity) (ablation)TextRIRO · ScenePairACC: 85.23 -> 4.37Hurt
Replace ControlNet injection with DiT-based OminiControl (ablation)TextRIRO · ScenePairACC (%): 85.23 -> 52.89Hurt
Use p_shuffle 0.7 instead of 0.5 (ablation)TextRIRO · ScenePairACC (%): 80.78 -> 82.42Improved
Use p_shuffle 0.7 instead of 1.0 (ablation)TextRIRO · ScenePairACC (%): 81.72 -> 82.42Improved
Pre-train text encoder with STR objective instead of glyph pre-training (ablation)TextRIRO · ScenePairACC / style fidelityImproved
Use SD 2.1 VAE instead of SD 1.5 (ablation)TextRIRO · ScenePairreconstruction quality: 28.33 -> 30.31Improved
Mix TextRIRO 1M data with real data for STRMAERec · Union14M-Benchmark avgaccuracy (%): 86.09 -> 87.98Improved
Scale TextRIRO data from 1M to 3MMAERec · Common Benchmarks avgaccuracy (%): 97.13 -> 96.98No clear effect
Scale TextRIRO data from 1M to 3MMAERec · Union14M-Benchmark avgaccuracy (%): 87.98 -> 88.00Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy