Layered Image Vectorization via Semantic Simplification

TL;DR

This paper introduces a progressive, layered image vectorization method that uses semantic simplification and a two-stage process to produce vectors with high fidelity, semantic alignment, and compact layered structure.

Contribution

The work presents a novel semantic simplification technique combined with a two-stage vectorization process for improved layered image vectorization.

Findings

High visual fidelity in vectorized images

Superior semantic alignment compared to existing methods

More compact layered representations

Abstract

This work presents a progressive image vectorization technique that reconstructs the raster image as layer-wise vectors from semantic-aligned macro structures to finer details. Our approach introduces a new image simplification method leveraging the feature-average effect in the Score Distillation Sampling mechanism, achieving effective visual abstraction from the detailed to coarse. Guided by the sequence of progressive simplified images, we propose a two-stage vectorization process of structural buildup and visual refinement, constructing the vectors in an organized and manageable manner. The resulting vectors are layered and well-aligned with the target image's explicit and implicit semantic structures. Our method demonstrates high performance across a wide range of images. Comparative analysis with existing vectorization methods highlights our technique's superiority in creating vectors with high visual fidelity, and more importantly, achieving higher semantic alignment and more compact layered representation. The project homepage is https://szuviz.github.io/layered_vectorization/.