Stylize — Benchmarks

Engine	Avg (ms)	Algorithm
ascii	15	Bitmap font, zero deps
halftone	28	Mono or CMYK rosette dots
dither	34	Floyd-Steinberg / Atkinson / Bayer
stipple	40	Jittered tone dots
xdog	54	XDoG edge detection
lowpoly	146	Edge-weighted Delaunay triangulation
pointillism	181	Seurat-style colored dots
crosshatch	220	Multi-angle tonal engraving
posterize	290	Color quantization + DoG ink
watercolor	378	Bilateral + pooling + granulation
fdog	385	Flow-based DoG, coherent lines
kuwahara	405	Anisotropic Kuwahara filter
pencil	493	Lu 2012 pencil sketch
cartoon	513	Abstraction + cel-shade + XDoG

Methodology

These numbers are averages across 3 synthetic fixtures (portrait 768×1024, landscape 1024×640, detail 800×800). Warmup runs excluded.

All processing is server-side on CPU — no GPU, no neural networks
Benchmarked on Apple M4 Pro (local dev). Production server (Linux x86, 2 vCPU) will be slower
Times vary with image size and complexity — larger images take proportionally longer
Kuwahara is the slowest engine (~1s) due to its per-pixel sectored filter kernel — optimization planned
Output is always PNG, which adds encoding time on top of the algorithm
Fixtures are deterministic (procedural, no network) — reproducible via bun run bench.ts

Real-photo quality comparison vs neural style transfer (NST) — we use classical algorithms, not AI. Different tradeoff: faster, deterministic, no model weights, but less "creative"
Production server times under load — single-threaded CPU, no queue yet
Memory usage per engine — not profiled, but all engines operate on raw pixel buffers