Lang-SVG: hierarchical image vectorization with language priors hits SOTA
clemson-computing · hf · 2026-11-21
A paper from Clemson, University of Copenhagen and eBay proposes Lang-SVG, tackling the problem that most vectorization methods optimize pixel fidelity but produce SVGs lacking human-perceived object-part hierarchies, making them hard to edit.
Key ideas:
- Explicit tree structure organizes SVG primitives with semantic labels and parent-child relations
- Granularity-controllable, open-vocabulary segmentation priors initialize object-part hierarchies
- MLLMs with SVG tree context assign semantic labels to unlabeled nodes
- Depth-adaptive simplification plus Tree-Guided Pruning and Merging removes redundant nodes
- L1-SAM-Prompting recovers missing regions for better fidelity and structural completeness
The paper also introduces a new evaluation protocol measuring structural quality and semantic part alignment; experiments show SOTA on both rendering fidelity and structural quality.
More from Multimodal
- Muse Code adds 3D model generation with printable textured output in CLI — charlieholtz · 2026-11-21
- Claude pitched 3 cinematic short films — from a circling dawn to one snowflake's fall — RileyRalmuto · 2026-11-21
- Alexandr Wang demos Muse: turn casual phone pics into sellable product shots — alexandr_wang · 2026-11-21
- ElevenMusic Adds Chaos Parameter to Control How Far Generated Music Strays — charis_ai · 2026-11-21
- Full workflow fixes Fable 5.2 video audio entirely inside Claude — AIwithGhotai · 2026-11-21
- NVIDIA ships Alpamayo 2 Edge, a 2B-parameter in-vehicle reasoning VLA model — NVIDIAAI · 2026-11-21