Enhance#
Week 4 asked whether evidence on a single output could tell two histories apart. This week: when someone shows a “before” and an “after” and calls the after a recovery of lost detail — what would actually verify that, and what would only make it look verified?
Three claims that are not the same claim#
Plausible. Looks like a coherent photograph. Weakest claim, easiest to satisfy.
Verified. The restored value at a named location matches a retained source. Strong — and requires exactly what a real restoration usually lacks: the original.
Uncertain, not fabricated. A flat, never-at-risk region matching tells you nothing about the method’s power.
This week’s pack, stated plainly#
An 8×8 synthetic tone chart, degraded by 2×2 block averaging, restored by bilinear interpolation.
Checked at six named locations: two uniform-block features matched (uninformative). Of four features with genuine variation: one matched, three did not.
One out of four is not a recovery rate — it’s what this pack’s own arithmetic produced.
A documented failure at higher stakes#
PULSE (Menon et al., CVPR 2020) searches a generative model’s space of plausible faces for one that downsamples to match a blurry input.
A pixelated photo of Barack Obama went in. White skin, blue eyes came out — verifiably wrong, because a real photo exists to check against.
The correct general lesson#
Not “generative restoration invents things,” as a blanket rule.
Plausibility was never evidence of correctness — and checking requires exactly the retained source a real restoration claim usually cannot supply.
Before Wednesday#
Download the pack. Read pixels.csv before restored.svg. Write down
which of the six features you expect to match before you check.
Workshop: Enhance
Text walkthrough#
For a reader who cannot use the deck view: three claims about a restoration are easily confused but not the same — plausible (looks like a coherent image, the weakest and easiest claim), verified (a restored value at a specific named location matches a retained source you can check against), and uncertain-not-fabricated (a region that was never at risk, such as a flat area, matching the source tells you nothing about the restoration method’s power, since flat areas survive almost any degradation and restoration). This week’s pack degrades an 8×8 synthetic tone chart by averaging each 2×2 block and restores it with ordinary bilinear interpolation, then checks six named locations against the retained original: two features sat in already-uniform blocks and matched uninformatively, while of four features with genuine internal variation, only one matched and three did not — one correct recovery out of four genuine attempts in this one specimen with this one method, not a general recovery rate. The same structure appears at much higher stakes in PULSE (Menon et al., CVPR 2020), a generative face-restoration method that, when run on a pixelated photo of Barack Obama, produced a plausible face with white skin and blue eyes — verifiably wrong because a real photograph exists to check against. The correct general lesson is not that generative restoration always invents detail, but that a plausible-looking output was never evidence of correctness on its own; checking requires exactly the retained source that a real restoration claim, unlike this pack’s, usually cannot supply.