Research map / Image generation / Autoregressive and latent generation
SVGBuilder: Component-Based Colored SVG Generation with Text-Guided Autoregressive Transformers
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 SVGBuilder, a text-guided autoregressive transformer that generates colored SVGs by assembling normalized, deduplicated path components, and releases the ColorSVG-100K dataset.
Problem
Generating high-quality colored SVGs from text is slow for optimization-based methods and hard for autoregressive token-level models because SVG code is long and paths are often invalid or incomplete.
Approach
Treat each SVG path as a reusable component: build a library by normalizing and merging similar paths (Jaccard similarity with union-find), then autoregressively predict component index, offsets, scale and RGB tokens conditioned on text and rasterized-image features, with a caption generator providing training captions.
Key results
- SVGBuilder achieves FID 15.93, the best overall, versus 20.28 for DiffSketcher and 41.03 for IconShop.
- Generation takes 1.36 s per SVG, a 604x speed-up over VectorFusion (about 13.69 min).
- Ablation: removing the Text Encoder training raises FID from 15.93 to 21.75; removing the Image Encoder changes FID least (to 16.90).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SVGBuilder instead of the strongest optimization-based baseline, DiffSketcher | ColorSVG-100K SVG generation · ColorSVG-100K | FID: 20.28 -> 15.93 | Helped |
| Use SVGBuilder instead of IconShop | ColorSVG-100K SVG generation · ColorSVG-100K | generation speed per SVG: 8.52 s (96.4 ×) -> 1.36 s (604.0 ×) | Helped |
| Add the Caption Generator (ablation) | SVGBuilder · ColorSVG-100K | FID: 17.40 -> 15.93 | Helped |
| Add the Image Encoder (ablation) | SVGBuilder · ColorSVG-100K | FID: 16.90 -> 15.93 | Helped |
| Train the Text Encoder (ablation) | SVGBuilder · ColorSVG-100K | FID: 21.75 -> 15.93 | Helped |
| Use color fills during inference (ablation) | SVGBuilder · ColorSVG-100K | FID: 17.39 -> 15.93 | Helped |
Surprising findings
- Removing the Image Encoder causes the smallest FID loss, suggesting the model does not effectively use its visual features.
- Optimization-based models do not degrade with SVG complexity while language-based models do.
Limitations
- Paths lack explicit semantics, so the model must infer them implicitly.
- Cross-entropy over number tokens may not capture geometric relationships.
- Component library limited to translation and scaling; rotation and reflection not used.
Experimental setup
Datasets: ColorSVG-100K Icon645
Benchmarks: ColorSVG-100K
Metrics: FID CLIPScore T2I CLIPScore I2I Human Preference Score Uniqueness Novelty generation speed per SVG FID by complexity level
Baselines: VectorFusion CLIPDraw DiffSketcher IconShop GPT-3.5 GPT-4
Builds on: IconShop DeepSVG StrokeNUWA CLIP GPT-2
Closest papers by meaning
- SVGDreamer: Text Guided SVG Generation with Diffusion Model (CVPR 2024 · similarity 0.91)
- Chat2SVG: Vector Graphics Generation with Large Language Models and Image Diffusion Models (CVPR 2025 · similarity 0.91)
- StarVector: Generating Scalable Vector Graphics Code from Images and Text (AAAI 2025 · similarity 0.91)
- OmniSVG: A Unified Scalable Vector Graphics Generation Model (NeurIPS 2025 · similarity 0.90)
- VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models (ICLR 2026 · similarity 0.88)
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.