Image Load Lab — WebP vs AVIF vs JPEG reproducible run

Run date: 2026-08-12
Fixture: ill-format-benchmark-source.png
Fixture origin: deterministic synthetic image created for this Image Load Lab run; no third-party media.
Dimensions: 1280 × 800 pixels, RGB, identical decoded source pixels and crop for every export.
SHA-256 (source PNG): 8ee761ff7a23c8ecccec1c2acf8db61dc975fca184e769169e01af4f35394da6

Environment
- Windows_NT 10.0.26200 (x64)
- sharp 0.34.5; libvips 8.17.3; mozjpeg 0826579; webp 1.6.0; heif 1.20.2; aom 3.13.1

Method
1. Generate one deterministic source that combines gradients, low-amplitude texture, 1 px diagonal lines, text edges and saturated color blocks.
2. Encode JPEG, WebP and AVIF ladders at quality inputs 35, 50, 65, 80, 92 without resizing or cropping.
3. Decode each file to RGB and compare it with the same source pixels.
4. Report RGB PSNR and 8×8-window luminance SSIM. These metrics are proxies, not perceptual verdicts.
5. Within each format, select the smallest tested file meeting PSNR ≥ 38 dB. If none meets it, select the highest-PSNR candidate.
6. Inspect the labeled identical crop in ill-format-benchmark-comparison.png.

Selected candidates
- JPEG Q92: 315137 bytes; PSNR 35.4652 dB; SSIM 0.943018; ill-format-benchmark-jpeg-q92.jpg
- WEBP Q92: 366050 bytes; PSNR 38.9323 dB; SSIM 0.975081; ill-format-benchmark-webp-q92.webp
- AVIF Q65: 186240 bytes; PSNR 38.2120 dB; SSIM 0.946278; ill-format-benchmark-avif-q65.avif

Limitations
- Encoder quality inputs are not equivalent across formats; the shared PSNR threshold is used to avoid treating equal quality numbers as equal quality.
- PSNR and SSIM do not capture every visible artifact or viewing condition.
- This synthetic fixture stresses gradients, fine detail, text edges and color blocks; it does not represent every photographic subject.
- Results belong to these exact encoder versions, settings, fixture and run date. Repeat with representative production images before changing a workflow.