Short for "designer in the loop".

Deterministic pipeline design for turning AI-generated product icon rasters into crisp, reusable SVGs.

This repo starts from a docs-first technical plan. The goal is to make icon vectorization measurable, automatable, and tunable with an AI controller without letting AI directly author final SVG geometry.

• Normalize noisy AI-generated icon rasters into traceable flat regions
• Vectorize with deterministic tools such as VTracer and Potrace
• Optimize and sanitize SVG output
• Snap geometry for small-size crispness
• Validate output with hard gates and a continuous score
• Use AI only to propose parameter changes between deterministic runs

• Overview
• Pipeline
• Scoring
• Debug Preview
• Roadmap

• Deterministic output beats opaque one-shot vector generation
• Small-size readability at 16, 20, and 24 px is the primary quality bar
• SVG simplicity matters: fewer paths, fewer nodes, fewer tiny artifacts
• Acceptance should be based on explicit thresholds and repeatable scoring

16, 20, 24, 32, 48, 64, 128, 256

• Ship: FinalScore >= 86 and all hard gates pass
• Needs tuning: 78 <= FinalScore < 86
• Reject: FinalScore < 78

• Raster cleanup and quantization
• VTracer for multi-color icon tracing
• Potrace for monochrome silhouettes and logos
• SVGO for SVG cleanup
• A scoring runner that renders candidate SVGs back to raster and evaluates them