Research map / Image generation / Autoregressive and latent generation
SuperSVG: Superpixel-based Scalable Vector Graphics 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 SuperSVG, a superpixel-based feed-forward image vectorization model with a coarse-stage and a coarse-path-guided refinement-stage transformer, trained with a novel Dynamic Path Warping loss to vectorize images into SVG paths quickly and accurately.
Problem
Image vectorization (raster to SVG) faces a trade-off: learning-based methods are fast but fail on complex images, while optimization-based methods reconstruct well but take very long.
Approach
Decompose the image into SLIC superpixels and vectorize each with a ViT-based coarse model, then train a refinement model that adds paths on top of the coarse canvas, using a DPW loss (a monotone, one-to-one DTW variant over path sequences) to avoid the vanishing-path local optimum.
Key results
- SuperSVG-B reaches PSNR 28.45 dB at 4,000 paths in 1.04 s, versus DiffVG 27.29 dB in 140.34 s and LIVE 26.80 dB in 300,000 s (optimization time).
- SuperSVG-F (10 s fine-tuning) achieves the best reconstruction at all path counts, e.g., PSNR 29.96 and MSE 0.0014 at 4,000 paths.
- Ablations show removing the refinement stage raises MSE from 0.0032 to 0.0041 and removing the DPW loss yields results similar to coarse-only output.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use superpixel segmentation instead of vectorizing the whole image without segmentation (ablation) | SuperSVG-B · ImageNet ablation, 1000 paths | MSE: 0.0083 -> 0.0032 | Helped |
| Use superpixel segmentation instead of 4 × 4 block division (ablation) | SuperSVG-B · ImageNet ablation, 1000 paths | MSE: 0.0038 -> 0.0032 | Helped |
| Add boundary loss (ablation) | Coarse-stage model · ImageNet ablation, coarse stage, 500 paths | MSE: 0.0534 -> 0.0045 | Helped |
| Add path efficiency loss (ablation) | Coarse-stage model · ImageNet ablation, coarse stage, 500 paths | MSE: 0.0063 -> 0.0045 | Helped |
| Add the refinement stage (ablation) | SuperSVG · ImageNet ablation, refinement stage | MSE: 0.0041 -> 0.0032 | Helped |
| Add DPW loss and coarse-path guidance (ablation) | SuperSVG refinement stage · ImageNet ablation, refinement stage | MSE: 0.0045 -> 0.0032 | Helped |
| Use DPW loss instead of L2 loss in path parameter space (ablation) | SuperSVG refinement stage · ImageNet ablation, refinement stage | MSE: 0.0045 -> 0.0032 | Helped |
| Fine-tune SuperSVG-B SVG parameters with L2 loss | SuperSVG-B · ImageNet test, 500 paths | MSE: 0.0044 -> 0.0026 | Helped |
| Fine-tune SuperSVG-B SVG parameters with L2 loss | SuperSVG-B · ImageNet test, 2000 paths | MSE: 0.0024 -> 0.0017 | Helped |
| Fine-tune SuperSVG-B SVG parameters with L2 loss | SuperSVG-B · ImageNet test, 4000 paths | MSE: 0.0019 -> 0.0014 | Helped |
| Use SuperSVG-B instead of LIVE | Image vectorization · ImageNet test, 500 paths | Time (s): 20,000 -> 0.31 | Helped |
| Use SuperSVG-F instead of DiffVG | Image vectorization · ImageNet test, 4000 paths | MSE: 0.0025 -> 0.0014 | Helped |
Surprising findings
- Refinement model trained with pixel L2 loss alone tends to shrink new paths to invisibility, a suboptimal local minimum.
- Replacing DPW with an L2 loss in path-parameter space gave results no better than omitting the path-space loss.
Limitations
- Refinement stage is trained with the coarse-stage model's output as pseudo ground truth, so refinement quality depends on the coarse model.
- Evaluation on ImageNet and EMOJIS only; LIVE evaluated on only 50 images due to cost.
Experimental setup
Datasets: ImageNet EMOJIS
Benchmarks: ImageNet test subset (1,000 images) EMOJIS
Metrics: MSE PSNR LPIPS SSIM inference time
Baselines: Potrace Adobe Illustrator Image Trace DiffVG LIVE Im2Vec Color Trace
Builds on: DiffVG AttnPainter DTW SoftDTW SLIC
Closest papers by meaning
- OmniSVG: A Unified Scalable Vector Graphics Generation Model (NeurIPS 2025 · similarity 0.81)
- StrokeNUWA—Tokenizing Strokes for Vector Graphic Synthesis (ICML 2024 · similarity 0.81)
- QuickRender: A Photorealistic Procedurally Generated Dataset with Applications to Super Resolution (Student Abstract) (AAAI 2024 · similarity 0.81)
- Vector Graphics Generation via Mutually Impulsed Dual-domain Diffusion (CVPR 2024 · similarity 0.81)
- Achieving Lightweight Super-Resolution for Real-Time Computer Graphics (AAAI 2025 · similarity 0.81)
- Empowering Vector Graphics with Consistently Arbitrary Viewing and View-dependent Visibility (CVPR 2025 · similarity 0.80)
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.