# Nxt: one-image export demonstration, 27 September 2026

This is an automated file-size experiment conducted for an Annemiek ter Haar
explainer, not a perceptual-quality benchmark or a ranking of codecs.
`input.png` is our own conceptual ImageGen illustration. All outputs use that
same RGB input at 1536 × 1024 pixels, with no resizing or metadata. The source
is a painted illustration, not a representative sample of photographs.

| Export | Exact bytes | Decimal kB |
| --- | ---: | ---: |
| JPEG | 248845 | 248.845 |
| WebP | 148696 | 148.696 |
| AVIF | 241351 | 241.351 |

All three use a numerical quality setting of 75. That setting does NOT match
visual quality across encoders. Smaller output does not by itself establish
better compression at equal quality. The decoded images retain the same pixel
dimensions, not necessarily the original pixel values.

Run `python3 site/web/research/image-formats-2026-09-27/measure.py` with Pillow
12.1.1. The measured build used Python 3.14.6, libjpeg-turbo 3.1.3, libwebp 1.6.0
and libavif 1.3.0. JPEG uses optimize=True and 4:2:0 subsampling; WebP method=6;
AVIF speed=6, 4:2:0 subsampling and one thread. See `measurements.json` for the
full settings, SHA-256 hashes and output dimensions. `sample.jpg`, `sample.webp`
and `sample.avif` are the actual encoded outputs. Each was decoded and visually
inspected; this inspection is not an objective quality score.

The article's native HTML chart uses bytes / 1000 with a common zero baseline
and a 300 kB maximum. Visible amounts round to one decimal. `hero.jpg` is a
separate quality-90 publishing export, not the JPEG entry in the experiment.

No third-party photograph or visual was reused. The input is AI-generated
editorial artwork in Nxt's matte gouache and paper-collage style.