Design Compass finds styles without names, running OpenCLIP in the browser

The project starts from a practical friction in image-generation prompting: advice often says to name a design style, but names like Future Funk, Art Deco, Vaporwave and Y2K are only useful once you already know them. Design Compass instead presents three candidates at a time, so users compare possibilities rather than classify an image into one correct style.

The implementation makes an interesting design choice about sliders. A single retro-versus-futuristic axis would force a tradeoff, so the app splits the sense of age and the sense of futurism into two independent sliders, allowing combinations such as retro-futurism. Internally, 512-dimensional OpenCLIP embeddings are reduced by PCA to 12 dimensions, and each style is modeled as a region with a mean and a Ledoit–Wolf covariance, ranked by Mahalanobis distance.

Search positions are recomputed from a fixed reference point on every search, so returning slider values returns the same results instead of drifting. For public release, the developer moved inference from a backend to the browser with ONNX Runtime Web and WebGPU, hosted on Cloudflare Pages with assets on Cloudflare R2 — free to run, though downloads still occur. A conversational mode remains listed as coming soon.