GlyphAtlas separates stroke detail from layout in scene text synthesis
Original title: GlyphAtlas: Disentangling Stroke Detail from Layout for Scene Text Synthesis
A glyph atlas supplies per-character stroke references so small, dense text is not limited by target-region resolution.
GlyphAtlas keeps a glyph image as a spatial anchor and adds an atlas that renders each character or word in its own cell. Atlas-anchored attention pulls stroke detail only from the matching cells to refine the coarse glyph. Training uses mask-weighted flow matching and a timestep-weighted text perceptual loss, plus random resolution augmentation for adjustable atlas budgets. The authors report accurate small and complex text, zero-shot transfer to unseen languages, and leading results on scenes with many text regions.
Unconfirmed: Performance claims are the authors' experimental results, not independently replicated here.
- Each character or word is encoded in a separate atlas cell.
- Random resolution augmentation supports adjustable atlas budgets at inference.
- Authors report zero-shot transfer to languages unseen in task training.
- The project page is fr0zencrane.github.io/project-glyphatlas.
From the original article
1 min read at huggingface.coScene text synthesis requires stroke-accurate characters rendered in a specified layout, as well as visual coherence with the surrounding scene. To achieve this, existing methods commonly pre-render the target text at its intended locations to form a glyph image that jointly specifies layout and glyph structure.
By Luozheng Qin, Changhai Zhou, Yiyi Zhou · 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.