Toolit background removal quality case study — 2026-10-12

Photos were NOT taken by Toolit. Each source's CC0 license was checked on the
experiment date. See sources.json for original file pages, credits, input
dimensions, preparation and SHA-256 hashes. CC0 does not imply endorsement
by the photographer or person pictured. The portrait is a pre-June-2017
Unsplash CC0 file; this is not a claim about all Unsplash photographs.

Files:
  *-input.png: full-frame prepared input (maximum long edge 1280px)
  *-balanced.png / *-detail.png: unretouched transparent browser PNG output
  *-comparison.webp: input / Balanced 512 / Detailed 768, navy compositing
  *-crop.webp: identical region from the same three files
  results.json: execution scope, runtime, model/code hashes and 8 outputs
  figures.json: crop rectangles, dimensions, display steps and figure hashes

Full-frame input files are not camera originals: public Wikimedia thumbnails
were normalized for orientation, stripped of metadata and saved as PNG.
Comparison images are resized, composited on #243247 and saved losslessly.
No sharpening, generative repair, color correction or manual masking was used.

Reproduce in the matching Toolit web checkout (Node 22.13+):
  npm ci
  npm run build
  node scripts/measure-background-removal-quality.mjs --prepare
  node scripts/measure-background-removal-quality.mjs --serve
  Open http://localhost:4186/ in Chrome and click the experiment button.
  Wait for all eight outputs, then stop the loopback server with Ctrl+C.
  node scripts/measure-background-removal-quality.mjs --figures

To reuse the exact archived inputs, skip --prepare (remote thumbnails may
change). Compare the local model/component/worker hashes to results.json.
The harness uses the actual built worker and current component helpers;
it fixes preferGpu=false (WASM), 512 and 768. It is not a complete UI test.
Inference sessions are released after each run. No analytics are emitted.
The harness saves outputs to this evidence directory on the same computer;
it is not a production upload endpoint. It binds to loopback only.

There are no ground-truth masks or representative accuracy percentages.
These four cases do not establish speed, mobile quality, WebGPU equivalence
or a competitor ranking. Future code/model/browser changes may alter output.