Research map / 3D vision / Face and shape
Towards High-fidelity Artistic Image Vectorization via Texture-Encapsulated Shape 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 a texture-encapsulated vector image representation where Bezier shape control points carry latent texture codes decoded by a shared MLP, optimized zero-shot per image for artistic image vectorization.
Problem
Existing image vectorization methods store a single color per shape and lose fine texture detail, while raster images couple geometry and texture and are hard to edit.
Approach
Attach trainable latent texture codes to Bezier control points (shape-anchored distribution) and decode coordinate-plus-interpolated-latent to RGB with a lightweight shared MLP; a geometric field built from Canny edges guides control point placement, avoiding sparse differentiable-rasterization gradients.
Key results
- On Emoji and Icon with 5 shapes, MSE 0.0007 (Emoji) and 0.0009 (Icon) versus 0.0019 and 0.0026 for LIVE.
- On Clipart with 128 shapes, MSE 0.0009 and SSIM 0.8803 versus 0.0090 MSE for LIVE.
- Against SIREN and grid-based latent codes, achieves MSE 0.0009 with 27,702 parameters versus 262,198 for Grid-/4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full texture-encapsulated shape representation instead of prior vectorization methods | Ours · Emoji and Icon | MSE: 0.0019 -> 0.0007; 0.0026 -> 0.0009 | Helped |
| Use the full texture-encapsulated shape representation instead of LIVE | Ours · Clipart | MSE: 0.0090 -> 0.0009; 0.0028 -> 0.0006 | Helped |
| Use shape-anchored latent codes instead of general implicit image representations | Ours · Clipart | MSE: 0.0058 -> 0.0009 | Helped |
| Add the geometric field loss (ablation) | Ours · Emoji, Icon, and Clipart | MSE: 0.0013 -> 0.0008; 0.0018 -> 0.0011; 0.0043 -> 0.0013 | Helped |
| Add the chamfer distance loss (ablation) | Ours with Lr + Lg · Clipart | MSE: 0.0013 -> 0.0009 | Helped |
| Replace IDW-Fusion with Sum-Fusion (ablation) | Ours · Emoji, Icon, and Clipart | MSE: 0.0007 -> 0.0013; 0.0009 -> 0.0012; 0.0009 -> 0.0016 | Hurt |
| Remove query coordinates from the texture feature input (ablation) | Ours · Emoji, Icon, and Clipart | MSE: 0.0007 -> 0.0015; 0.0009 -> 0.0020; 0.0009 -> 0.0042 | Hurt |
| Increase the number of shapes (ablation) | Ours · Clipart | MSE | Helped |
| Increase latent code dimension (ablation) | Ours · Clipart | MSE | No clear effect |
Surprising findings
- LIVE reaches comparable Clipart reconstruction when given roughly half the shapes of the proposed method at 256 shapes, per the text.
Limitations
- Requires per-image zero-shot optimization rather than a feed-forward model.
- Number of shapes must be set manually (e.g., 5, 128, or 256).
Experimental setup
Datasets: Emoji Icon Clipart dataset (200 images, introduced)
Benchmarks: Emoji Icon Clipart dataset
Metrics: MSE LPIPS SSIM parameter count
Baselines: DiffVG LIVE NPA SIREN LIIF-style grid latent codes
Builds on: DiffVG LIVE LIIF SIREN
Closest papers by meaning
- Optimize & Reduce: A Top-Down Approach for Image Vectorization (AAAI 2024 · similarity 0.87)
- Vector Grimoire: Codebook-based Shape Generation under Raster Image Supervision (ICML 2025 · similarity 0.85)
- Easy-editable Image Vectorization with Multi-layer Multi-scale Distributed Visual Feature Embedding (CVPR 2025 · similarity 0.85)
- Learning to Manipulate Artistic Images (AAAI 2024 · similarity 0.84)
- Less is More: Efficient Image Vectorization with Adaptive Parameterization (CVPR 2025 · similarity 0.83)
- Training-free Spatially Grounded Geometric Shape Encoding (NeurIPS 2026 · 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.