Research map / Image generation / Autoregressive and latent generation
Layered Image Vectorization via Semantic Simplification
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 progressive layered image vectorization method that uses SDS-based feature-average simplification to generate coarse-to-fine target images, then optimizes vectors in structure and visual refinement stages.
Problem
Image vectorization usually reconstructs a raster image as overly complex, intricate vector shapes that are hard to manage and poorly aligned with semantic objects.
Approach
Use Score Distillation Sampling with the conditional text prompt set to empty (increasing unconditional noise share) to produce progressively simplified images via the feature-average effect; guide layered vector optimization with masks from these simplified images in a two-stage structure-then-fidelity process.
Key results
- Vector Compactness of 73.8% vs 65.9% for the next best (SGLIVE) on 100 test images, with lower MSE and LPIPS than all four baselines across primitive counts.
- Higher CLIP similarity between captions of the input and rasterized vectors, especially with few vector paths.
- Ablation shows the simplified-image sequence recovers implicit semantic parts (e.g., grassland, whole body) missed without it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a sequence of simplified images rather than only the input image (ablation) | Layered Image Vectorization via Semantic Simplification · Image vectorization | semantic alignment | Helped |
| Replace SDS-based simplification with bilateral filtering (ablation) | Layered Image Vectorization via Semantic Simplification · Image vectorization | boundary preservation | Hurt |
| Replace SDS-based simplification with Gaussian filtering (ablation) | Layered Image Vectorization via Semantic Simplification · Image vectorization | boundary preservation | Hurt |
| Replace SDS-based simplification with superpixel-based simplification (ablation) | Layered Image Vectorization via Semantic Simplification · Image vectorization | boundary preservation | Hurt |
| Use SDS-based simplification instead of conventional simplification methods (ablation) | Layered Image Vectorization via Semantic Simplification · Image vectorization | semantic recovery | Helped |
| Use the proposed layered vectorization method rather than SGLIVE | Image vectorization · 100-image testing dataset | VeC (%): 65.9 -> 73.8 | Helped |
| Use the proposed method rather than DiffVG, LIVE, O&R, or SGLIVE | Image vectorization · 100-image testing dataset | MSE | Helped |
| Use the proposed method rather than other compared methods | Image vectorization · Sampled path counts | CLIP similarity score | Helped |
Limitations
- Relies on semantic segmentation of simplified images, so segmentation errors propagate to vector layers
- Evaluation on a relatively small 100-image test set
Experimental setup
Datasets: 100-image test set of realistic photos, clipart, and emojis
Benchmarks: 100-image test set (photos, clipart, emojis)
Metrics: MSE LPIPS Vector Compactness (VeC) CLIP score
Baselines: DiffVG LIVE Optimize & Reduce SGLIVE Bilateral filtering Gaussian filtering Superpixel simplification
Builds on: Score Distillation Sampling DiffVG LIVE Classifier-Free Guidance
Closest papers by meaning
- Less is More: Efficient Image Vectorization with Adaptive Parameterization (CVPR 2025 · similarity 0.87)
- Optimize & Reduce: A Top-Down Approach for Image Vectorization (AAAI 2024 · similarity 0.86)
- Channel-wise Vector Quantization (NeurIPS 2026 · similarity 0.84)
- Easy-editable Image Vectorization with Multi-layer Multi-scale Distributed Visual Feature Embedding (CVPR 2025 · similarity 0.84)
- VectorArk: Learning Practical Image Vectorization with Rounded Polygon Representation (CVPR 2026 · similarity 0.84)
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · 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.