AI upscaling to 4K pays off where content is worst: restoring degraded SD buys ~20× more quality per Wh than polishing clean HD — and re-processing pristine 4K buys nothing at the highest energy cost

The result, in one sentence

Across a twelve-rung degradation ladder (two contents × six quality rungs, all upscaled/restored to 4K by the same AI pipeline under identical measurement), the energy cost is set almost entirely by the input resolution (~45 Wh per 45 s clip for 4K input, ~12–15 Wh for HD/SD input), while the quality gain rises monotonically as source quality falls — so the energy-to-quality sweet spot sits squarely on the worst content: heavily degraded SD gained +2.1 NR-VQA points for 14.8 Wh, clean HD gained +0.23 for 11.9 Wh, and re-processing pristine 4K gained nothing (±0.0) for 44–46 Wh.

The ladder

Quality columns: CompressedVQA-HDR (no-reference — "how good does it look?", the same metric /enhance-run reports) scored on input and output; VMAF-4K (full-reference — "how faithful to the master?"), possible here uniquely because every rung descends from a known reference cut.

| input (45 s, 60 fps) | CompressedVQA-HDR in | out | ΔVQA | ΔE (Wh) | ΔVQA/Wh | VMAF-4K vs master (FR) | |---|---|---|---|---|---|---| | bbb_sd_dirty | 5.45 | 7.56 | +2.11 | 14.85 | 0.142 | 47.4 | | meridian_sd_dirty | 6.04 | 7.64 | +1.60 | 14.55 | 0.110 | 67.6 | | meridian_sd_clean | 7.63 | 9.01 | +1.38 | 14.58 | 0.095 | 81.0 | | bbb_sd_clean | 7.83 | 9.02 | +1.19 | 14.49 | 0.082 | 67.7 | | meridian_hd_dirty | 7.95 | 8.46 | +0.51 | 11.89 | 0.043 | 82.5 | | bbb_hd_dirty | 8.10 | 8.62 | +0.52 | 12.12 | 0.043 | 72.0 | | bbb_4k_dirty | 8.63 | 8.93 | +0.30 | 45.30 | 0.007 | 74.1 | | meridian_4k_dirty | 9.16 | 9.44 | +0.28 | 46.21 | 0.006 | 85.5 | | bbb_hd_clean | 9.30 | 9.53 | +0.23 | 11.93 | 0.019 | 88.9 | | meridian_hd_clean | 9.49 | 9.74 | +0.25 | 12.03 | 0.021 | 94.3 | | bbb_ref_4k | 9.58 | 9.63 | +0.05 | 43.54 | 0.001 | 93.2 | | meridian_ref_4k | 9.86 | 9.84 | −0.02 | 44.56 | ~0 | 97.6 | | 2005 UGC anchor (21 s, 15 fps) | 5.37 | 7.20 | +1.83 | 1.73 | — | n/a (no master) | | Night UGC anchor (44 s, 21 fps) | 7.89 | 8.53 | +0.64 | 4.25 | — | n/a (no master) |

All fourteen runs returned 🟢 confidence on the dual-meter instrument (recalibrated same day; GPU variance 0.91%).

Four learnings

1. Energy is a property of the input pixels, not the achieved quality. The pipeline drew ~390 W regardless of rung; what varied was duration-of-work per frame: 4K input ≈ 45 Wh, SD input (×4.5 scale) ≈ 14.6 Wh, HD input (×2) ≈ 12 Wh per identical 45 s/60 fps clip. The ×4.5 SD path costs slightly more than ×2 HD despite fewer input pixels — scale factor modulates, input resolution dominates. Cost is therefore predictable before running: it does not depend on how degraded the content is.

2. Quality gain is a property of the input's badness. With cost fixed per rung, ΔVQA rises monotonically as source quality falls — from ±0.0 (pristine) to +2.1 (worst). The per-Wh spread across the ladder exceeds 100× (0.142 vs ~0.001). The economic reading: an operator with a restoration budget should spend it on the worst material first, and never on already-good 4K, where the same energy buys literally nothing.

3. Synthetic and cinematic content behave the same — on the NR axis. The BBB (synthetic animation) and Meridian (cinematic camera) curves overlap within ~0.2 ΔVQA at every matched rung, and their energy columns are near-identical. The difference appears on the full-reference axis: Meridian restores more faithfully at every rung (round-trip VMAF 97.6 vs 93.2 on the pristine refs; 67.6 vs 47.4 on the worst rung — a 20-point gap) — consistent with a restoration model trained predominantly on camera content (though BBB's higher measured spatial/temporal complexity — SI ~33 vs ~13 — is an alternative explanation: harder content is harder to restore faithfully; see caveats). Perceived improvement is content-agnostic; fidelity recovery is not.

4. Looking better is not being restored. The NR scores say enhanced SD-clean output is "near-pristine" (9.01–9.02, within 0.6 of the actual masters); the full-reference scores say those same outputs sit at VMAF 67.7–81.0 against their masters — and the worst rung's output, NR-scored a respectable 7.56, round-trips at VMAF 47.4. The enhancement manufactures plausible detail that a no-reference metric (and plausibly a viewer) rewards, but the original information is not recovered. OWL's standing position — perceptual quality of super-resolution is for the viewer to judge — survives contact with this data: the energy buys palatability, not fidelity, and the two metrics quantify the difference for the first time on this rig.

The real-UGC anchors validate the synthetic ladder

The genuinely bad 2005 camera clip (NR 5.37 → 7.20, +1.83) lands almost exactly on the synthetic curve between the two sd_dirty rungs (5.45 → 7.56 and 6.04 → 7.64); the mid-quality night clip (7.89 → +0.64) falls between the hd_dirty (+0.51) and sd_clean (+1.19) rungs, as its 720p VFR character predicts. The artificial degradation recipe — the most attackable part of this experiment's design — produces quality-gain behaviour indistinguishable from real-world degradation at matching quality levels.

How it was measured

Twelve 45-second fixtures generated from the best practical public copies of Big Buck Bunny (Sunflower 2160p60) and Meridian (Netflix Open Content) by documented, reproducible recipes — frozen with checksums and per-file NR scores in the library manifest. Each fixture (plus two real UGC clips) ran through the standard /enhance-run harness: focus mode, per-run baseline, 1 Hz wall-power polling on the primary meter of a dual daisy-chained Tapo P110 pair, NR quality (CompressedVQA-HDR) scored on input and output as a terminal pass outside the energy window. Full-reference PSNR/SSIM/VMAF (vmaf_4k model, 4× temporal subsample) computed post-hoc against each content's reference cut — possible here, uniquely among enhancement runs, because every rung descends from a known master. Interactive chart, fixture downloads, and recipes: /enhance-run/ladder.

What this finding does not measure

• Other enhancement pipelines, models, or targets (one vendor pipeline, balanced model, 4K target only).
• HD as an output target (today's data says 4K-target energy ≈ 4× HD-target at the same input — see the S45 UGC runs — but this sweep held the target fixed).
• Whether viewers prefer the enhanced output (no subjective study; the NR metric is a learned proxy).

The traditional-upscaler comparison (two runs)

On the SD rungs, a plain Lanczos upscale to 4K (NVENC-encoded at the same 35 Mbps) cost 0.63–0.64 Wh vs the AI's 14.5 Wh (~23×). The AI delivered 2–5× the quality gain: sd_clean — Lanczos +0.41 NR vs AI +1.19; sd_dirty — Lanczos +0.37 vs AI +2.11. One honest nuance: on these synthetic rungs Lanczos slightly improved the NR score, whereas on the real 2005 UGC clip (earlier same-day run) it lowered it (−0.17) — whether cheap upscaling helps or harms appears to depend on the artifact character of the source. Both compare runs 🟢 (enhance/a5c4ce9c, enhance/69fc5a1d).