GlyphAtlas: Disentangling Stroke Detail from Layout for Scene Text Synthesis
Abstract
GlyphAtlas uses a dedicated glyph atlas and atlas-anchored attention to provide fine-grained stroke references for scene text synthesis, improving stroke accuracy and zero-shot transfer.
Scene 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. However, this representation ties the resolution of each glyph to its target region, limiting fine-grained stroke guidance for small, densely packed, or structurally complex text. To tackle this, we propose GlyphAtlas, a scene text synthesis framework that retains the glyph image as a spatial anchor and supplies fine-grained stroke references through a dedicated glyph atlas. The atlas renders and encodes each character or word in a separate cell, decoupling reference resolution from target-region size. We further introduce atlas-anchored attention, which lets each text region retrieve stroke detail only from its corresponding atlas cells to refine the coarse glyph structure provided by the glyph image. This coarse-to-fine conditioning simplifies stroke rendering while preserving explicit layout correspondence. We optimize GlyphAtlas with a mask-weighted flow matching objective and a timestep-weighted text perceptual loss for writing accuracy and background consistency. Random resolution augmentation of atlas cells supports adjustable atlas budgets at inference. Extensive experiments demonstrate that GlyphAtlas achieves stroke-accurate synthesis of small, dense, and structurally complex text, zero-shot transfer to languages unseen during task-specific training, and leading performance in scenes with many text regions. project page: https://fr0zencrane.github.io/project-glyphatlas/
Community
🚀 Excited to introduce GlyphAtlas, a scene text synthesis model built on FLUX.1-Fill-dev—fully open-sourced with code, weights, and a 1.5M dataset!
⚠️ Why existing methods struggle: Standard approaches rely on a scene-aligned glyph image to specify both layout and character shape. Because glyph resolution is tied to region size, small text, dense multi-line layouts, and complex scripts get squeezed into a handful of pixels, resulting in missing, merged, or distorted strokes.
🧠 Our approach: GlyphAtlas decouples where text goes from how strokes are drawn. A scene-aligned glyph image anchors layout, while a resolution-independent glyph atlas—encoded directly via the backbone's native VAE without extra encoders—supplies crisp stroke references. Atlas-anchored attention then binds each cell back to its target region for accurate coarse-to-fine synthesis.
🌍 Zero-shot & Open Source: Trained solely on Chinese and English, GlyphAtlas-Bi generalizes zero-shot to Japanese, Korean, and Arabic, handling everything from curved storefronts and posters to dense book covers. Alongside our code and weights, we also release GlyphAtlas-Data-1.5M (1.5 million scene text images) to support the community!
🌐 Project page: https://fr0zencrane.github.io/project-glyphatlas/
🤗 Model checkpoint: https://huggingface.co/Fr0zencr4nE/GlyphAtlas
📚 Dataset: https://huggingface.co/datasets/Fr0zencr4nE/GlyphAtlas-Data-1.5M
💻 Code: https://github.com/Fr0zenCrane/GlyphAtlas
Try it out and let us know what you think, feedback and questions are very welcome! 🙌
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Enhanced Video Text Editing with Trajectory-Aligned Glyph Rendering (2026)
- Planning and Rendering in Concert: DeepFusion of Autoregressive Layouts and Diffusion for Visual Text Generation (2026)
- Efficient Scene-Text Refinement via Pixel-Space Grouped Image Generation (2026)
- Real In, Real Out: What If We Only Use Real Data for Scene Text Editing? (2026)
- Four-Step Text-Bearing Image Generation via Distillation with Local Response Regularization and Teacher-Guided Training (2026)
- RefCompose: Multi-Reference Image Generation via LoRA-Conditioned Diffusion (2026)
- Decompose Radicals, Then Reward: Fine-Grained Inspection for Accurate Chinese Text Rendering (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
Get this paper in your agent:
hf papers read 2611.20192 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper