Research map / Image generation / Editing and style transfer
IF-Font: Ideographic Description Sequence-Following Font 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 IF-Font, a few-shot font generator that replaces the source-glyph content input with an Ideographic Description Sequence (IDS) and autoregressively predicts VQGAN tokens of the target glyph conditioned on reference-glyph tokens.
Problem
Few-shot font generation methods rely on disentangling content and style from a source glyph, which is hard for complex CJK glyphs and leads to missing strokes, blur and artifacts when source and target fonts differ greatly.
Approach
Drop the source glyph and use the character's IDS as content control; quantize glyphs with a pretrained VQGAN and model target tokens autoregressively given the IDS and reference-glyph style features (IHA, SSA with global/local aggregation, and a contrastive SCE loss).
Key results
- On UFSC with 3 reference glyphs, IF-Font reports FID 6.84 and L1 0.1258 versus FontDiffuser's 3.70 FID and 0.1248 L1, i.e. FontDiffuser has better FID in several settings while IF-Font leads on L1/LPIPS/SSIM.
- User study: IF-Font preferred in 22.78% of UFSC 3-shot judgments versus 18.32% for FontDiffuser (second best).
- Ablation: adding IHA, SSA and SCE to the baseline improves FID from 8.27 to 8.49 (full) with L1 dropping 0.1632 to 0.1597; SSA global branch removal yields FID 7.98.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IF-Font instead of FontDiffuser | few-shot font generation · UFSC 1-shot | FID: 3.9969 -> 6.7695 | Hurt |
| use IF-Font instead of FontDiffuser | few-shot font generation · UFSC 3-shot | FID: 3.6989 -> 6.8359 | Hurt |
| use IF-Font instead of FontDiffuser | few-shot font generation · UFSC 8-shot | FID: 4.1017 -> 6.7383 | Hurt |
| use IF-Font instead of FontDiffuser | few-shot font generation · UFUC 1-shot | FID: 8.2524 -> 8.4844 | Hurt |
| use IF-Font instead of FontDiffuser | few-shot font generation · UFUC 3-shot | FID: 7.6444 -> 8.4922 | Hurt |
| use IF-Font instead of FontDiffuser | few-shot font generation · UFUC 8-shot | FID: 8.9166 -> 8.3203 | Helped |
| add IHA, SSA, and SCE to the baseline (ablation) | IF-Font · UFSC and UFUC, 3-shot | FID: 8.2656 -> 8.4922 | Hurt |
| add IHA (ablation) | IF-Font baseline · UFSC and UFUC, 3-shot | FID: 8.2656 -> 8.3750 | Hurt |
| add SSA to the IHA variant (ablation) | IF-Font · UFSC and UFUC, 3-shot | FID: 8.3750 -> 7.5391 | Helped |
| add SCE to the IHA+SSA variant (ablation) | IF-Font · UFSC and UFUC, 3-shot | FID: 7.5391 -> 8.4922 | Hurt |
| use stroke-level instead of component-level IDS (ablation) | IF-Font · UFSC and UFUC, 3-shot | FID: 8.4922 -> 8.4297 | Helped |
| use mixed component and stroke IDS granularity (ablation) | IF-Font · UFSC and UFUC, 3-shot | FID: 8.4922 -> 8.5234 | Hurt |
Surprising findings
- FontDiffuser achieves lower FID than IF-Font in several 1-shot and 3-shot settings despite IF-Font leading on L1, LPIPS and SSIM.
- Stroke-level IDS granularity degrades performance relative to component-level IDS, due to conflicts between IDS sequences.
Limitations
- Struggles with fancy, irregular, flat/narrow, heavily curved or calligraphic fonts; VQGAN reconstruction errors cap quality.
- IDS is imperfect: some characters share identical stroke-level IDS and spatial relations are underspecified.
- Autoregressive inference is slow and attention scales quadratically with sequence length.
Experimental setup
Datasets: self-collected Chinese font rendering set (424 training fonts, 3,300 characters; 40 unseen test fonts) IDS decomposition table (babelstone IDS.TXT, simplified to 12 IDCs)
Benchmarks: UFSC (unseen fonts, seen characters) UFUC (unseen fonts, unseen characters)
Metrics: FID L1 LPIPS RMSE SSIM user preference (%)
Baselines: CG-GAN LF-Font FS-Font CF-Font VQ-Font NTF FontDiffuser
Builds on: VQGAN nanoGPT MoCo
Closest papers by meaning
- VQ-FONT: Few-Shot Font Generation with Structure-Aware Enhancement and Quantization (AAAI 2024 · similarity 0.89)
- Generate Like Experts: Multi-Stage Font Generation by Incorporating Font Transfer Process into Diffusion Models (CVPR 2024 · similarity 0.88)
- VecGlypher: Unified Vector Glyph Generation with Language Models (CVPR 2026 · similarity 0.87)
- DeepCalliFont: Few-Shot Chinese Calligraphy Font Synthesis by Integrating Dual-Modality Generative Models (AAAI 2024 · similarity 0.87)
- FontDiffuser: One-Shot Font Generation via Denoising Diffusion with Multi-Scale Content Aggregation and Style Contrastive Learning (AAAI 2024 · similarity 0.85)
- Decoupling Layout from Glyph in Online Chinese Handwriting Generation (ICLR 2025 · similarity 0.84)
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.