Keeping the Red Red: Colour Loss, Perceptual Quality Targets and Deployment Constraints in a Self-Hosted Storefront Image Pipeline

Abstract

Self-hosted storefronts compress and resize every product photo they serve, and the tools that do it treat quality as a preset and ignore the colour profile inside the file. We study one WooCommerce store on shared hosting end to end: an image pipeline built on the three core WordPress filters that also govern the browser-side processing introduced in WordPress 7.1, a measurement method for the colour loss that occurs when a wide-gamut profile is dropped, and a calibration of encoder quality against structural similarity per product category. The pipeline cut a 93-image library from 126.6 MB to 87.3 MB in WebP with no visible loss. In that library 13 photos (14%) carry a Display P3 profile, far above the 0.4% share reported for the web; on a GD host, which drops profiles, two of them shift every pixel by a mean ΔE76 of 6 to 7, well past the just-noticeable distance. A crawl of 68 live modest-fashion stores (26 with a public catalogue, 2,419 images) found wide-gamut profiles on only 0.5% of served images and no loss in the wild: Shopify converts its resized copies to sRGB and the one self-hosted store with such uploads keeps the profile on resize. The common defect there is size, with 19% of full-size product images narrower than 900 px. The calibration shows that the spread of structural similarity inside a category is as wide as the gap between categories, so a per-image perceptual target is the only setting that is both small and safe. Testing both upload paths of WordPress 7.1 exposed a sequencing defect that orphans the untouched upload when an image is both converted and scaled, and we give the correction. Finally we analyse the constraints a browser-side pipeline inherits: cross-origin isolation, which limits it to Chromium today; the cost of a perceptual metric in WebAssembly; and the absence of a reliable colour transform on shared hosting. The measurement tools, the crawler and the data are public.