LexiAlign: A Diffusion Model Text Alignment and Refinement Method Based on Local Regeneration

Abstract Despite the recent advances of diffusion models in high-quality image synthesis, generating visually accurate textual content remains a persistent bottleneck—commonly manifesting as misspellings, glyph distortions, and semantic drift. We introduce LexiAlign, a language-guided local diffusion refinement framework that directly targets these issues through three tightly coupled components: robust optical character recognition (OCR)-based text extraction, language model-driven semantic correction, and high-fidelity local inpainting via masked diffusion. Unlike prior approaches that retrain large diffusion backbones or overwrite entire regions, LexiAlign performs character-level targeted repair while preserving surrounding visual context and style. To support systematic training and evaluation, we construct SynOCRText, a 120k-sample benchmark covering 8 languages, over 20 fonts, diverse layouts, and fine-grained error masks. On SynOCRText, LexiAlign achieves 88.4% OCR accuracy (+6.3% over the best baseline), Contrastive Language-Image Pretraining (CLIP) Score of 0.852 (+0.023), Peak Signal-to-Noise Ratio (PSNR) of 30.92\,dB (+2.51\,dB), and Structural Similarity Index Measure (SSIM) of 0.893 (+0.020). These results establish LexiAlign as a plug-and-play, domain-agnostic solution for reliable visual text alignment, offering both quantitative superiority and practical deployability for creative design, advertising, and multimodal content generation.