Research map / Image generation / Autoregressive and latent generation
Optimize & Reduce: A Top-Down Approach for Image Vectorization
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 Optimize & Reduce (O&R), a top-down DiffVG-based vectorization method that clusters colors with DBSCAN, iteratively optimizes Bézier shapes and halves the shape count via a removal score, plus a geometric loss against self/cross intersections, and contributes a five-dataset vectorization benchmark.
Problem
Raster-to-vector image conversion (vectorization) that yields compact, editable SVGs with few shapes while remaining fast and domain-agnostic; existing optimization methods are slow or produce redundant shapes, and learning-based methods generalize poorly.
Approach
Top-down vectorization: initialize many Bézier shapes from DBSCAN color clusters and connected components, optimize all shapes jointly with DiffVG, then repeatedly score shapes by the reconstruction loss increase when each is removed and drop the lowest-ranked ones (exponential schedule 256→128→64), with a differentiable geometric loss penalizing self- and cross-intersections.
Key results
- On the new emoji dataset O&R reaches MSE comparable to LIVE while being roughly two orders of magnitude faster (LIVE averages about 30 minutes for a 64-shape image).
- O&R outperforms DiffVG with random initialization in MSE across all five datasets for a fixed shape count; it also beats DiffVG with DBSCAN initialization except on NFT-Apes.
- The paper reports a speedup of over x10 over LIVE in runtime.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Optimize & Reduce instead of DiffVG with random initialization | DiffVG vectorization · All five datasets | MSE | Helped |
| Use Optimize & Reduce instead of DiffVG with DBSCAN initialization | DiffVG vectorization · All datasets except NFT-Apes | MSE | Helped |
| Use DiffVG with DBSCAN initialization instead of random initialization (ablation) | DiffVG · All datasets | MSE | Helped |
| Use DBSCAN initialization with DiffVG (ablation) | DiffVG · NFT-Apes | MSE | Helped |
| Add the O&R optimization and reduction scheme after DBSCAN initialization (ablation) | DiffVG with DBSCAN initialization · Low shape counts | MSE | Helped |
| Add 32 shapes after reducing to 32 shapes (ablation) | Optimize & Reduce · Images with intricate details or prominent textures | reconstruction quality | Helped |
| Fix LIVE's image-loading bug (ablation) | LIVE · New EMOJI, 8 shapes | MSE: 5.3 -> 1.5 | Helped |
| Use Optimize & Reduce instead of LIVE | Image vectorization · Vectorization benchmark | runtime: × 10 | Helped |
Surprising findings
- After fixing a pre-processing bug in LIVE's image loading, LIVE's MSE at 8 shapes dropped from 5.3 to 1.5 gray levels, changing the comparison with the original reported results.
- DiffVG with DBSCAN initialization outperforms O&R on the NFT-Apes dataset, where the clustering initialization introduces a strong inductive bias.
Limitations
- Single-image optimization method; slower than feed-forward learning-based vectorizers and not trained on datasets.
- Results on complex images with fine textures can be coarse at low shape counts, requiring an add operation to recover details.
- Default baseline and LIVE comparisons depend on re-implementation and bug fixes made by the authors.
Experimental setup
Datasets: NotoEmoji (old 2015 version) NotoEmoji (new 2022 version) Free-SVG NFT-Apes Midjourney text-to-image images
Benchmarks: NotoEmoji old NotoEmoji new Free-SVG NFT-Apes Midjourney
Metrics: MSE LPIPS runtime
Baselines: DiffVG LIVE LIVE-ES AutoTrace Potrace VectorFusion
Builds on: DiffVG LIVE DBSCAN CLIPasso
Closest papers by meaning
- Less is More: Efficient Image Vectorization with Adaptive Parameterization (CVPR 2025 · similarity 0.90)
- VectorArk: Learning Practical Image Vectorization with Rounded Polygon Representation (CVPR 2026 · similarity 0.87)
- Towards High-fidelity Artistic Image Vectorization via Texture-Encapsulated Shape Parameterization (CVPR 2024 · similarity 0.87)
- Layered Image Vectorization via Semantic Simplification (CVPR 2025 · similarity 0.86)
- VFIG: Vectorizing Complex Figures in SVG with Vision-Language Models (NeurIPS 2026 · similarity 0.86)
- Bézier Splatting for Fast and Differentiable Vector Graphics Rendering (NeurIPS 2025 · similarity 0.85)
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.