Research map / Image generation / Text-to-image diffusion
Decoupling Layout from Glyph in Online Chinese Handwriting Generation
Read the paperCheck your idea against 108,967 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.
Proposes a hierarchical framework for stylized online Chinese handwritten text line generation, pairing an autoregressive in-context LSTM layout planner with a multi-scale contrastive style-encoder diffusion character synthesizer.
Problem
Generating complete online Chinese handwritten text lines with specified content and a reference writer's style, which prior work (mostly single-character generation) has left largely unexplored.
Approach
Factorizes text line generation into layout planning (autoregressive LSTM planner that uses a style reference prefix as in-context prompt to predict character bounding boxes) and per-character diffusion synthesis (1D U-Net denoiser with character embedding dictionary and multi-scale style encoder trained with layer-wise contrastive loss, fused via cross-attention).
Key results
- On individual characters, the method reaches DTW 0.932 and Style Score 0.918, the best among purely data-driven methods (DeepImitator, DiffWriter), while SDT still scores higher on DTW and Content Score.
- Text lines generated by the method achieve Accurate Rate 0.857 and Correct Rate 0.852 (GT: 0.962 and 0.958) with Style Score 0.910.
- In-context layout planning yields lower geometric feature deviations than unconditional and Gaussian layout sampling on nearly all features.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multi-scale contrastive learning at all feature layers (ablation) | Stylized diffusion character synthesizer · ICDAR-2013 character generation | Style Score: 0.875 -> 0.918 | Helped |
| add multi-scale contrastive learning at all feature layers (ablation) | Stylized diffusion character synthesizer · ICDAR-2013 character generation | Content Score: 0.893 -> 0.891 | No clear effect |
| use contrastive learning at scale 3 rather than no contrastive loss (ablation) | Stylized diffusion character synthesizer · ICDAR-2013 character generation | Style Score: 0.875 -> 0.900 | Helped |
| use contrastive learning at scale 3 rather than no contrastive loss (ablation) | Stylized diffusion character synthesizer · ICDAR-2013 character generation | Content Score: 0.893 -> 0.887 | No clear effect |
| use in-context reference-conditioned layout planning rather than unconditional layout generation | Layout planner · Text line layout generation | ∇2: 0.1328 -> 0.1229 | Helped |
| use in-context reference-conditioned layout planning rather than unconditional layout generation | Layout planner · Text line layout generation | ∇5: 0.1605 -> 0.1293 | Helped |
Surprising findings
- Decoupling layout planning from character synthesis and training them separately was found to work better than joint training, since joint training with unnormalized character sizes produced unstable generation.
- The style-transfer-based SDT outperforms the proposed method on DTW (0.880) and Content Score (0.970) in the individual character comparison.
Limitations
- Each character is generated independently, so the method struggles to mimic styles with extensive cursive connections between characters.
- Reports no error bars; results averaged over three test groups only.
- Code is promised upon acceptance but not released.
Experimental setup
Datasets: CASIA-OLHWDB 1.0-1.2 CASIA-OLHWDB 2.0-2.2 ICDAR-2013 competition database
Benchmarks: ICDAR-2013 (individual character test set) CASIA-OLHWDB 2.0-2.2 test writers (text lines)
Metrics: DTW Content Score Style Score binary geometric layout features (nabla 1-8) Accurate Rate Correct Rate
Baselines: DeepImitator DiffWriter FontRNN WriteLikeU SDT Gaussian layout sampling Unconditional layout planner
Builds on: DDPM DeepImitator DiffWriter SketchRNN
Closest papers by meaning
- DeepCalliFont: Few-Shot Chinese Calligraphy Font Synthesis by Integrating Dual-Modality Generative Models (AAAI 2024 · similarity 0.87)
- VQ-FONT: Few-Shot Font Generation with Structure-Aware Enhancement and Quantization (AAAI 2024 · similarity 0.86)
- GlyphMastero: A Glyph Encoder for High-Fidelity Scene Text Editing (CVPR 2025 · similarity 0.86)
- A Multi-Style Chinese Characters Writing Intelligent Tool Based on Small-scale Training Data (AAAI 2025 · similarity 0.85)
- VecGlypher: Unified Vector Glyph Generation with Language Models (CVPR 2026 · similarity 0.85)
- Layout-Agnostic Scene Text Image Synthesis with Diffusion Models (CVPR 2024 · similarity 0.85)
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.