Go to ICLR 2026 Conference homepageLayerSVG: Layer-wise Semantic Editable Scalable Vector Graphics Synthesis
Keywords: Scalable Vector Graphics, Layer-wise Vectorization
TL;DR: We propose LayerSVG, a novel model that converts complex raster images into editable, semantically layered SVGs with efficiency, addressing the long-standing challenge of generating hierarchically editable SVGs in prior methods.
Abstract: Scalable Vector Graphics (SVG) is a lightweight and editable image format. Converting complex raster images into semantically layered and editable SVGs presents a longstanding challenge. Existing vectorization methods primarily focus on holistic image conversion, producing a single, uneditable SVG, but neglecting SVG layering that is crucial for SVG editing. Although some approaches attempt simple layer extraction, they are often limited to basic icons or individual strokes. To address these limitations, we propose LayerSVG, a novel method capable of top-down, semantic layer-wise vectorization of complex raster images. Our method employs a layer-elimination strategy to progressively decompose layers, extract semantic objects and inpaint obscured regions from top to bottom. For robustly determining object occlusion relationships, we design a robust three-stage judgment mechanism, ensuring high accuracy and automated extraction. Furthermore, for optimal stroke allocation across multiple layers, we propose an adaptive path allocation mechanism, which considers layer area and complexity to efficiently utilize the finite SVG path budget. Extensive experiments, encompassing fidelity tests and diverse editing tasks, and comprehensive computational resource analysis, demonstrate that LayerSVG not only achieves powerful reconstruction and versatile editable layers, but also runs efficiently. This fills a critical gap in the field of semantically editable SVG conversion from raster images.
Primary Area: learning on graphs and other geometries & topologies
Submission Number: 6122
Loading