VectorNova open-sources NovaClip, a clip-style model optimized for icon and glyph retrieval

Tech · 3 min read

NovaClip is a compact embedding model that encodes SVGs, path metadata, and bitmap renderings into low-dimensional vectors tuned for perceptual similarity. VectorNova trained NovaClip on millions of icons, glyphs, and UI assets to capture fine-grained visual semantics like stroke weight, corner treatment, and metaphors.

The team released reference integrations for Figma, Sketch, and a server-side npm package, enabling instant semantic search, duplicate detection, and automated tagging pipelines. NovaClip supports mixed queries: text plus example icon to refine search results based on both semantics and visual style.

By open-sourcing NovaClip, VectorNova hopes to standardize how design tools index and retrieve vector assets, reducing reliance on general-purpose image embeddings that often miss vector-specific features. Community contributors are already proposing downstream plugins for icon normalization and cross-platform mapping.