Skip to content

Performance & Benchmarks: dusrel-image ​

Budgets (hard gates in CI, via size-limit) ​

See TRD.md section 5. A budget increase requires an ADR.

Benchmark method ​

  • Runner: benchmarks/run.mjs (Node, against the built packages). The Benchmarks workflow runs it weekly and on demand and uploads the report as an artifact; benchmarks/results.md is the committed record of the latest run.
  • Corpus: three deterministic synthetic images generated inside the runner from a seeded PRNG — a photo-like gradient with per-pixel noise (384×288), flat UI-like blocks (800×600) and an alpha pattern (256×256). Licence-clean by construction: the generator is the source. A corpus of real photographs is future work; until then, treat the numbers as regression tracking, not marketing claims.
  • Times are medians (5 runs after 1 warmup; AVIF 3, metadata 25), single-threaded WASM exactly as shipped. Node timings differ from browser timings; the Playwright suite is the browser correctness gate, this benchmark is the size/time regression gate.

Measured numbers (benchmarks/run.mjs, 2026-11-09, Node 24) ​

Same photo-like input (384×288, 442,368 raw RGBA bytes) through each encoder:

OperationOut (bytes)Median (ms)
AVIF encode q50, speed 97,28531.1
JPEG encode q80 (MozJPEG)14,85934.6
WebP encode q80 (libwebp)23,85659.1
PNG encode (lossless)213,46050.5
OxiPNG optimise of that PNG, level 2191,632122.1
JPEG decode—2.7
Resize lanczos3 to 50%—4.0

Full table (14 operations, incl. quantize, placeholders, metadata, SVG): benchmarks/results.md. Upstream research figures quoted during planning (MozJPEG ~53% median saving at q75 on real photos, etc.) were superseded by these first-party measurements on 2026-11-09.

Time targets ​

  • Interactive operations (resize/crop/rotate + JPEG/WebP encode of a 12 MP photo) should complete in under ~2 s on a mid-range laptop, under ~5 s on a mid-range phone, in the worker, without blocking the UI.
  • AVIF is exempt from time targets but must always show progress and support cancellation.