Research map / Image generation / Autoregressive and latent generation
StrokeNUWA—Tokenizing Strokes for Vector Graphic Synthesis
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 StrokeNUWA, which tokenizes SVG strokes into discrete stroke tokens via a residual-VQ compressor (VQ-Stroke) so an encoder-decoder LLM (Flan-T5) can generate vector graphics from text quickly.
Problem
LLM-based visual synthesis relies on raster grid tokens from VQ-VAE/VQ-GAN, which break semantic structure and are slow; optimization-based SVG generation is also very slow. The paper targets text-guided SVG generation with LLMs.
Approach
Represent SVG paths as sequences of basic commands (M, L, C) packed into a 9-dim matrix, then compress them into semantically meaningful stroke tokens with a residual vector quantizer (VQ-Stroke); an encoder-decoder LLM (Flan-T5 3B, frozen encoder, fine-tuned decoder) generates these tokens from text, and a post-processing SVG Fixer repairs syntax/connectivity.
Key results
- StrokeNUWA (PI) reaches FID 6.513, CLIPScore 17.994 and HPS 16.801, beating VectorFusion (7.754 / 17.539 / 15.901) and Iconshop (17.828 / 8.402 / 8.234).
- Generation takes about 19.1 seconds per SVG versus about 30 minutes for LIVE/VectorFusion-style optimization, reported as up to 94x speedup.
- VQ-Stroke with compression rate 4 reaches a compression ratio of 6.9% of the tokenized SVG code length while keeping FID and CLIPScore comparable to rate 2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace direct SVG-code tokens with VQ-Stroke stroke tokens | Flan-T5 SVG generation · SVG generation | FID: 17.828 -> 6.513 | Helped |
| replace direct SVG-code tokens with VQ-Stroke stroke tokens | Flan-T5 SVG generation · SVG generation | CLIPScore: 8.402 -> 17.994 | Helped |
| replace direct SVG-code tokens with VQ-Stroke stroke tokens | Flan-T5 SVG generation · SVG generation | HPS: 8.234 -> 16.801 | Helped |
| use Path Interpolation instead of Path Clipping (ablation) | SVGNUWA · SVG generation | FID: 6.607 -> 6.513 | Helped |
| use Path Interpolation instead of Path Clipping (ablation) | SVGNUWA · SVG generation | CLIPScore: 17.852 -> 17.994 | Helped |
| use Path Interpolation instead of Path Clipping (ablation) | SVGNUWA · SVG generation | HPS: 16.134 -> 16.801 | Helped |
| add Path Interpolation rather than Path Clipping (ablation) | VQ-Stroke C-2 reconstruction · SVG reconstruction | EDIT: 1,096.313 -> 1,315.137 | Hurt |
| add Path Interpolation rather than Path Clipping (ablation) | VQ-Stroke C-2 reconstruction · SVG reconstruction | FID: 3.751 -> 3.518 | Helped |
| increase compression rate from C-2 to C-4 (ablation) | VQ-Stroke with Path Interpolation · SVG reconstruction | Compression ratio: 13.780% -> 6.890% | Helped |
| increase compression rate from C-2 to C-4 (ablation) | VQ-Stroke with Path Interpolation · SVG reconstruction | CLIPScore: 20.290 -> 17.192 | Hurt |
| reduce codebook size from 4096 to 2048 while keeping dim 512 (ablation) | VQ-Stroke · SVG reconstruction | FID: 3.518 -> 5.702 | Hurt |
| increase codebook dimension from 512 to 1024 while keeping codebook size 4096 (ablation) | VQ-Stroke · SVG reconstruction | FID: 3.518 -> 3.901 | Hurt |
Surprising findings
- GPT-4 produces only simple SVGs and often yields graphics incongruent with the text prompt.
- Without the SVG Fixer, VQ-Stroke outputs fail to conform to SVG syntax.
Limitations
- Evaluated only on monochromatic black-and-white icons from FIGR-8-SVG, not real-world images.
- Relies on a post-processing SVG Fixer, since raw generations often violate SVG syntax or path connectivity.
- Human evaluation and qualitative comparisons are limited in scale (50 prompts).
Experimental setup
Datasets: FIGR-8-SVG
Benchmarks: FIGR-8-SVG test split (2,000 instances; 500 sampled for baseline comparison)
Metrics: FID CLIPScore HPS EDIT Recall generation speed compression ratio
Baselines: VectorFusion LIVE (with Stable Diffusion) Iconshop GPT-4
Builds on: VQ-VAE IconShop DeepSVG VectorFusion
Closest papers by meaning
- Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.85)
- Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.85)
- SVGDreamer: Text Guided SVG Generation with Diffusion Model (CVPR 2024 · similarity 0.84)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.84)
- Optimize & Reduce: A Top-Down Approach for Image Vectorization (AAAI 2024 · similarity 0.84)
- UniFlow: A Unified Pixel Flow Tokenizer for Visual Understanding and Generation (ICLR 2026 · 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.