Research map / Image generation / Editing and style transfer
Less is More: Efficient Image Vectorization with Adaptive Parameterization
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 AdaVec, an optimization-based image vectorization method that decomposes a raster image into pure-colored layers (SAM + SLIC + DBSCAN), simplifies Bezier control points adaptively, and refines shapes and colors with differentiable rendering (DiffVG).
Problem
Existing image vectorization methods use fixed numbers of paths and control points, which are redundant for simple images and insufficient for complex ones, and optimization-based methods are slow.
Approach
Adapt the number of paths and Bezier control points to the shape complexity of each layer: decompose the image into perceptually aligned layers, prune redundant control points via a chamfer-distance-fitted simplification, then jointly optimize shape and color with differentiable rendering and a geometric intersection loss.
Key results
- On Noto Emoji, AdaVec (N=39) reaches PSNR 35.89 and SSIM 0.982 with 1220 parameters in 44.9 s, versus LIVE (N=256) PSNR 34.53 at 9342.2 s.
- On Fluent Emoji, AdaVec (N=26) achieves PSNR 33.28 with 1012 parameters in 50.5 s, versus O&R (N=256) PSNR 33.10 at 253.5 s.
- On Iconfont, AdaVec (N=215) achieves PSNR 28.03 and LPIPS 0.0504, outperforming O&R (N=256) PSNR 25.13 while using fewer parameters than 7168 for O&R-256 (4598 vs 7168).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use adaptive path and control-point parameterization | AdaVec · Noto Emoji | MSE: 0.00036 -> 0.00032 | Helped |
| use adaptive path and control-point parameterization | AdaVec · Fluent Emoji | MSE: 0.00065 -> 0.00047 | Helped |
| use adaptive path and control-point parameterization | AdaVec · Iconfont | MSE: 0.00127 -> 0.00102 | Helped |
| use adaptive path and control-point parameterization | AdaVec · Noto Emoji | generation time (s): 9342.2 -> 44.9 | Helped |
| use adaptive path and control-point parameterization | AdaVec · Fluent Emoji | generation time (s): 6783.0 -> 50.5 | Helped |
| use adaptive path and control-point parameterization | AdaVec · Iconfont | generation time (s): 19731.9 -> 164.2 | Helped |
| add superpixel segmentation (ablation) | AdaVec multi-layer decomposition · Noto Emoji | MSE: 0.00085 -> 0.00032 | Helped |
| add optimized decoupling (ablation) | AdaVec control-point simplification · Noto Emoji | MSE: 0.00114 -> 0.00032 | Helped |
| add superpixel segmentation (ablation) | AdaVec multi-layer decomposition · Fluent Emoji | MSE: 0.00152 -> 0.00047 | Helped |
| add optimized decoupling (ablation) | AdaVec control-point simplification · Fluent Emoji | MSE: 0.00068 -> 0.00047 | Helped |
| add superpixel segmentation (ablation) | AdaVec multi-layer decomposition · Iconfont | MSE: 0.00356 -> 0.00102 | Helped |
| add optimized decoupling (ablation) | AdaVec control-point simplification · Iconfont | MSE: 0.00258 -> 0.00102 | Helped |
Surprising findings
- Fixed-path baselines with more paths can produce redundancy on simple images, so more paths do not always yield better reconstruction.
Limitations
- Precise reconstruction of extremely complex gradient images is not yet achieved, as shown in failure cases.
- Radial gradients and stop-color quantity are not dynamically selected.
- Relies on hand-tuned thresholds (MIoU, alpha, delta, lambda) set to fixed values.
Experimental setup
Datasets: Noto Emoji Fluent Emoji Iconfont
Benchmarks: Noto Emoji Fluent Emoji Iconfont SVGEdit-Bench
Metrics: MSE LPIPS PSNR SSIM generation time number of parameters number of paths
Baselines: O&R SGLIVE LIVE
Builds on: DiffVG O&R LIVE VTracer
Closest papers by meaning
- Optimize & Reduce: A Top-Down Approach for Image Vectorization (AAAI 2024 · similarity 0.90)
- Layered Image Vectorization via Semantic Simplification (CVPR 2025 · similarity 0.87)
- VectorArk: Learning Practical Image Vectorization with Rounded Polygon Representation (CVPR 2026 · similarity 0.84)
- Towards High-fidelity Artistic Image Vectorization via Texture-Encapsulated Shape Parameterization (CVPR 2024 · similarity 0.83)
- Easy-editable Image Vectorization with Multi-layer Multi-scale Distributed Visual Feature Embedding (CVPR 2025 · similarity 0.83)
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · similarity 0.82)
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.