Research map / Image generation / Editing and style transfer
Easy-editable Image Vectorization with Multi-layer Multi-scale Distributed Visual Feature Embedding
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 a multi-layer, multi-scale parameterized image representation that embeds texture features at hierarchical proxy geometry (Bézier boundary points plus Triwild-refined triangle vertices) per semantic layer, enabling accurate natural-image vectorization and easy shape/texture editing.
Problem
Existing parameterized/vectorized image representations embed texture along semantic boundaries, so they cannot model internal textures of natural images and lose content consistency under editing.
Approach
Decompose the image into SAM-based semantic layers ordered by depth, fit each layer with one adaptive Bézier shape, triangulate its interior with Triwild and gradient-driven multi-scale refinement, and store texture features at both boundary and interior proxy vertices, decoded by a small MLP with inverse-distance and barycentric interpolation.
Key results
- On ImageNet, the method reaches MSE 0.37e-3 and PSNR 34.32 versus 1.40e-3 and 29.96 for the best prior vectorizer S-SVG (4000 shapes).
- On Clipart, MSE 0.17e-3 and PSNR 37.70 versus 0.55e-3 and 32.59 for Chen et al.
- Compared with implicit baselines on ImageNet, it uses 48,919 parameters and 2,106 codes versus 113,411 parameters for Chen-512, with better MSE (0.37 vs 1.56).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multi-layer multi-scale texture embedding | Ours · ImageNet | MSE (×10−3): 1.56 -> 0.37 | Helped |
| use multi-layer multi-scale texture embedding | Ours · Clipart | MSE (×10−3): 0.55 -> 0.17 | Helped |
| add internal proxy triangles to edge control points (ablation) | Ours · ImageNet | MSE (×10−3): 1.13 -> 0.54 | Helped |
| add multi-scale internal triangles instead of single-scale triangles (ablation) | Ours · ImageNet | MSE (×10−3): 0.54 -> 0.37 | Helped |
| use layer-wise Bézier fitting instead of triangulating the entire image (ablation) | Ours · ImageNet | MSE (×10−3): 0.65 -> 0.37 | Helped |
| add position encoding (ablation) | Ours · ImageNet | MSE (×10−3): 0.93 -> 0.37 | Helped |
| use barycentric interpolation instead of inverse-distance weighting within triangles (ablation) | Ours · ImageNet | MSE (×10−3): 0.43 -> 0.37 | Helped |
Limitations
- Per-image optimization takes minutes per image rather than feed-forward inference.
- Editing evaluation is qualitative and compared only against a modified version of Chen et al.
Experimental setup
Datasets: ImageNet Icon Emoji Clipart
Benchmarks: ImageNet Clipart Emoji Icon
Metrics: MSE PSNR LPIPS SSIM
Baselines: DiffVG Im2Vec LIVE O&R SuperSVG Chen et al. SIREN LIIF
Builds on: Chen et al. shape-anchored distributed texture representation LIIF Triwild SAM Depth Anything
Closest papers by meaning
- Towards High-fidelity Artistic Image Vectorization via Texture-Encapsulated Shape Parameterization (CVPR 2024 · similarity 0.85)
- Layered Image Vectorization via Semantic Simplification (CVPR 2025 · similarity 0.84)
- Improving Editability in Image Generation with Layer-wise Memory (CVPR 2025 · similarity 0.83)
- Less is More: Efficient Image Vectorization with Adaptive Parameterization (CVPR 2025 · similarity 0.83)
- Generative Image Layer Decomposition with Visual Effects (CVPR 2025 · similarity 0.83)
- Optimize & Reduce: A Top-Down Approach for Image Vectorization (AAAI 2024 · similarity 0.83)
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.