TraceForge

Started by benchmarking every open-source vectorization engine available. Potrace excelled at clean geometric shapes (logos, icons) while VTracer handled photographic complexity better. Rather than picking one, I built a dual-engine architecture letting users choose the right tool per asset. The key insight was that GPU-accelerated neural upscaling before vectorization dramatically improves output quality—feeding a 4x upscaled image to Potrace produces cleaner paths than running Potrace on the original. Built the pipeline on FastAPI with async processing, WebSocket progress streaming for long batch operations, and an SVGO post-processing stage that strips metadata and optimizes path data. Added 14 vectorization presets tuned for different asset types: logos, icons, illustrations, technical drawings, and photographs.

TraceForge ships as a self-hosted web application with a React frontend and FastAPI backend. The GPU pipeline handles neural upscaling via CUDA-accelerated models, then routes to either Potrace or VTracer based on user selection or automatic detection. WebSocket connections stream real-time progress for batch operations processing hundreds of assets. The SVGO optimization stage runs 12 plugins that reduce SVG file sizes by 40-60% without visual degradation. Currently processing 2,000+ conversions monthly with zero manual intervention. The entire pipeline runs on a single RTX 3080 with 8-second average processing time per asset—down from 45 minutes of manual work.