The blog

Scanning guides, restoration stories, and a little of the science behind bringing old photos back. Every figure in the engineering posts comes from a real benchmark or production log — and when something did not work, the post says so.

FEATUREDTransparent PNGs break AI upscalers: the dark halo bugA −94-luma trench three pixels wide, traced to the canvas premultiplied-alpha round-trip — and the O(n) flood fill that removes 98% of it.AI upscaling vs bicubic: a 28-image benchmarkKodak-24 plus four historical scans through five reconstruction methods: the perception–distortion tradeoff, measured live on our production path.Are the colors real? Chroma-only AI colorizationWhy colourisation here can never change detail — only tint it — and the 115-photo evaluation of the rule that decides when it runs.How much should AI enhancement upscale? A bounded designThe planning rule that makes whole-image super-resolution fast and predictable regardless of upload size — now with seam, padding, and cost curves.Why a restored photo can be a larger file than the scanSix times the pixels plus a bloated browser PNG encoder: where the bytes come from and the 45% we got back.Can AI repair tears and scratches? Feasibility numbersThe measurements that took damage repair from proof of concept to a shipping opt-in beta, controls included.A face detector once restored a toddler's pantsThe false positive that motivated a geometric sanity gate, the exact rules — and the measured detection floor, Solvay 1927 included.Why AI-restored faces look like video game charactersEight rendering variants benchmarked across 35 faces to find why restored faces look plastic — then the fix quantified with face-recognition metrics.How to restore old photos, step by stepThe whole journey — scanning, repairing, upscaling, and archiving — with real before/after results along the way.