GlyphAtlas disentangles stroke detail from layout for accurate scene text synthesis

Luozheng Qin · hf · 2026-11-17

Researchers from Xiamen University and Fudan University present GlyphAtlas for scene text synthesis. A dedicated glyph atlas with atlas-anchored attention disentangles stroke detail from layout, providing fine-grained stroke references during generation.

Qualitative results show stroke-accurate text blending seamlessly into dense, high-density scenes such as movie posters, departure boards, and storefront signage, with improved stroke accuracy and zero-shot transfer.

Original post →

More from Multimodal

Multimodal channel →