Reproduce “Compositional SVG Generation via VLM-Driven Hierarchical Semantic Parsing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sehwan Park, Taehoon Kim, Geonhee Han, Dohyun Kim, Seung Wook Kim, Paul Hongsuck Seo
A VLM-guided pipeline converts raster illustrations into hierarchical SVGs whose groups correspond to semantic objects rather than arbitrary color regions. By combining recursive parsing, segmentation, amodal completion, inpainting, and vectorization, the method improves object-level editing: it achieves a post-edit rendering error of 0.0210 versus 0.0400 for optimally grouped VTracer outputs, while retaining competitive whole-image fidelity.
Key insights
- Semantic SVG generation is better framed as scene parsing before vectorization than as path generation followed by grouping.
- Amodal completion is important for editability: recovering hidden geometry prevents holes and improves the rendering outcome after object removal or movement.
- Structural quality and rendering fidelity are distinct objectives; VTracer can achieve lower whole-image error while producing less useful object-level organization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.