Week 10 Workshop: Enhance

Does a convincing restoration match the original?

5 May 2027

Does a convincing restoration match the original?

Learning target. Distinguish plausible detail, verified agreement and remaining ambiguity.

Where this comes from#

Week 4 asked whether the evidence on a single output could separate two histories, and left Output 7 in the archive with its parentage only partly established. That week said the question would come back here. This is where it comes back, from the other direction: not “which history made this?” but “does putting the detail back get you the original?” Week 4 could not rule out a route; this week asks whether a restoration that looks right actually agrees with the source, cell by cell.

Before the workshop#

Download the week 10 pack: three images — source.svg, degraded.svg, restored.svg — built from a single 8x8 synthetic tone chart, not a photograph. degraded.svg replaces every 2x2 block of the source with its average tone; restored.svg upsamples that degraded grid with bilinear interpolation, a standard, named algorithm, not a mystery process. pixels.csv has every one of the 64 cells for all three images. Do not open reveal/ yet.

Readings.

1. Menon, Damian, Hu, Ravi & Rudin, “PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models,” CVPR 2020, freely available at arxiv.org/abs/2003.03808. Read §1–3 for what the method actually does: it searches a generative model’s own space of plausible faces for one that, when downsampled, matches the low-resolution input — it does not reconstruct the original high-resolution pixels.
2. Quach, K. (2020, June 24). “Once again, racial biases show up in AI image databases, this time turning Barack Obama white.” The Register. theregister.com/2020/06/24/ai_image_tool. A documented case: PULSE, given a pixelated photo of Barack Obama, produced a fabricated face with white skin and blue eyes — verifiably wrong, because a real photo of Obama exists to check against. Read for the article’s own caveat: PULSE was never designed to recover an input person’s true appearance, only to find a plausible generated face consistent with the blur, so the correct description of its failure here is narrower than “the AI invented a person.”

Check — 5 minutes#

Feature f4 in this week’s pack sits inside a block whose four source cells are not all equal — real information was lost when that block was averaged — yet the restored value at f4 exactly matches the retained source. Does that match show bilinear interpolation recovered real detail at f4? State your answer and why in one sentence.

Receive — 15 minutes#

Confirm you can state, in your own words, the difference between a value that is plausible (it looks like it could belong in the image), a value that is verified (it has been checked against retained source evidence and matches), and a value that is uncertain (the check available cannot tell you anything, because nothing was actually at risk of being lost there).

Investigate — 40 minutes#

Examine the three images (10 minutes). Compare source.svg, degraded.svg and restored.svg (or their text equivalents in labelled-features.txt). Note that restored.svg looks smoother than degraded.svg — visual plausibility is not yet a claim about correctness.

Complete the feature table (20 minutes). For each of the six named features in feature-table.csv, record: is the source cell inside a block that had internal variation (i.e., could restoration have lost real information there), and does the restored value match the source value? Classify each feature as supported (matches, and the block had internal variation to recover from), unsupported (does not match), or uncertain (the block was already uniform, so a match proves nothing about restoration — there was nothing to recover).

Test a claimed recovery and state a limitation (10 minutes). Pick the one feature your table calls supported. State exactly what would have to be true for that same algorithm to be trusted on a different specimen without a retained source to check against — and note that this pack cannot supply that answer, because it has only ever been tested on this one specimen.

Evaluate — 30 minutes#

Write your feature table (id, location, judgement, one line of reasoning per row) before opening reveal/. Then answer: if you had access to degraded.svg and restored.svg only — no retained source — which of your six judgements could you actually have made?

Deposit — 30 minutes#

Open reveal/analysis.md (ask your tutor for the link once your predictions are written) and compare it against your own table. Deposit your feature table, labelled supported/unsupported/uncertain, into the Image collection.

Accessible route#

labelled-features.txt gives the same six features as plain text — location, source tone, degraded tone, restored tone — with no image required.

Assessment#

The Lossless Argument requires evaluating a preservation or restoration claim using retained evidence, weighted for how well you separate verified agreement from plausible appearance. See the assessment brief.