JPEG vs WebP in Pixo: a reproducible benchmark
This benchmark runs Pixo's current smart-compression engine against matched JPEG and WebP inputs generated from four deterministic canvases. The result is a transparent engineering reference—not a universal claim that one format always wins.
Method and corpus
The public-domain synthetic corpus contains gradient/photo-like color, flat illustration, text-interface, and deterministic noise classes at 640×480. Each canvas is encoded to JPEG and WebP at browser quality 0.92, then Pixo processes it at quality 90, a 1920-pixel maximum dimension, and the never-ship-a-bigger-file rule. No user or third-party image is included.
Aggregate result
Across 8 cases in Chromium 149.0.7827.55, Pixo reduced the JPEG inputs by 0.53% in aggregate and the WebP inputs by 11.36%. The never-bigger rule kept already-efficient outputs unchanged instead of publishing larger files.
How image class changed the gap
The matched WebP output was 77.23% smaller than JPEG for flat-illustration, while the narrowest observed gap was 10.43% for noise-texture. These values describe this corpus and method only; they are generated from results.json rather than copied into page content.
Reproduce it
Run pnpm benchmark:images:check to rebuild the unpacked compression runtime and rerun all eight cases. The check fails if an output moves outside the committed 2% tolerance. Run pnpm benchmark:images only when intentionally updating the versioned artifact after reviewing codec, browser, corpus, or settings changes.
Limits and interpretation
These are byte comparisons for matched synthetic canvases, not a blinded visual-quality study and not a representative sample of every photo or graphic. The JPEG and WebP source files were already encoded before Pixo processed them, so both source encoding and recompression affect the result. Browser, codec, settings, dimensions, transparency, and content can change the outcome; test representative production images before choosing a default format.