ohms.studio

writing  ·  .0004  ·  2026

we tested real-esrgan on archival film and rejected it

the upscale looked sharper and was less faithful, which is a bad trade for an archive.

what we were doing

remastering archival lecture footage for a foundation that holds the estate. the source is decades old, transferred more than once, and the grain is part of the record rather than noise on top of it.

why the upscale failed

a super resolution model does not recover detail that was never recorded. it invents plausible detail and renders it convincingly. on a face that means fabricated skin texture and edges that were never in the negative. it looks sharper on a still and it is less faithful to what was filmed, which for an archive is the wrong direction entirely.

what we did instead

a colour grade. warm, soft and rich rather than clean and clinical, matched to how the material would have looked projected. nothing added, nothing sharpened, nothing invented. the standard on the whole project is that we only ever remove, never add.

the general rule

generative restoration is fine when the artefact is the point and wrong when the record is. an archive is a record. if a viewer could mistake invented detail for evidence, you have damaged the thing you were hired to preserve.

next → how i film a website, and the three bugs that took a week