| Takeaway | Detail |
|---|---|
| Only accept sharpening if it delivers a median PSNR gain of at least 0.5 dB. | The thesis sets the bar at a 0.5 dB median PSNR gain across a CRF 23/28/32 ladder. |
| Reject sharpening if static-block flicker goes up by more than 10%. | The reader rule caps the increase in the peak flicker index at 10%. |
| Require a block-crawl count of 0 out of 3. | The reader rule says to use temporal-consistency-constrained sharpening only when the block-crawl count is 0 out of 3. |
| Run the 3-clip CRF ladder before making a call. | The reader rule calls for a 3-clip CRF ladder plus all three acceptance checks. |
The sharpening acceptance thresholds — the 0.5 dB median PSNR floor, the 10% flicker ceiling, and the 0-out-of-3 block-crawl requirement — are summed up in the takeaways table above and applied in the gate sections below.
That answer flips when temporal consistency fails: if the peak flicker increase is over 10% or the block-crawl count isn’t 0 out of 3, temporal-consistency-constrained sharpening loses out to deblocking/denoising.

How sharpening turns blocks into flicker
Compression leaves behind an 8×8 block grid whose edge energy shifts with the quantizer, so a per-frame sharpening kernel treats every block boundary like real detail. It’s the same mechanism that makes cross-scale consistency constraints useful: genuine structure holds up across spatial, temporal, and frequency scales, while artifacts tend to stay stuck in a single scale, as the EmergentAn overview of cross-scale consistency constraints explains. When a sharpening pass boosts high-frequency energy in one frame without checking neighboring frames, it turns a static coding grid into a time-varying edge pattern. That is the point where blocks start to flicker.
Arc consistency gives the cleanest pruning rule. The KTIML constraint guide states that once a constraint graph is node-consistent, unary constraints can be removed because they are satisfied. By analogy, prune sharpening strengths that violate temporal consistency before ranking by PSNR: a candidate that cannot maintain its edge response across adjacent frames is removed from the pool rather than averaged into a score. In practice, check each candidate against the same block-edge locations in the preceding and following frames, then keep only those whose response survives that pairwise test. This keeps the ranking from rewarding a single-frame gain that the eye reads as crawl.
Local consistency constrained adaptive neighbor embedding adds the second half of the mechanism. The ResearchGate paper on that method describes a local consistency constraint that explicitly regularizes linear reconstruction and adaptively generates as many candidates as possible for high-resolution output. Translated to compressed video, the reconstruction is the sharpened block, and the candidates are the allowed neighbor patterns around it. Keep only sharpening candidates that preserve local block structure across adjacent frames. If a candidate changes the slope of a block edge relative to its neighborsneighbors from one frame to the next, it fails the local-consistency check even if its single-frame contrast looks better.
The two checks work at different levels. Arc consistency removes candidates that break the pairwise temporal rule; local-consistency regularization removes candidates that break the spatial neighborhood rule inside the reconstruction itself. Together they explain why unconstrained per-frame sharpening amplifies block-grid flicker: the kernel has no term that ties an 8×8 boundary to its counterpart in the next frame, and no term that keeps the neighbor pattern stable while it raises edge energy. The result is a grid that appears to pulse as the quantizer reallocates bits.
For the reader’s rule, this is the reason to inspect the flicker index and block-crawl count before accepting any PSNR gain. A sharpening pass that wins on still-frame measurement but fails the temporal pair test is precisely the case the mechanism predicts. Treat the block grid as a constraint graph, not as detail, and the choice becomes a consistency problem first and a sharpness problem second.

What the evidence actually supports
The evidence here is narrower than the decision it is being asked to support. Emergentmind, updated May 27, 2026, defines cross-scale consistency constraints as principles that enforce agreement and invariance between multi-scale observations in spatial, temporal, and frequency domains. That is a design constraint: if a sharpening pass is going to be called temporally consistent, it needs to preserve structure across those scales. It isn't, on its own, a PSNR claim, and it doesn't give you a video-quality delta you can plug straight into a gate.
MatchAttention (arXiv) is the one source that actually reports a directional result for a consistency-constrained objective. Its snippet says the Consistency-Constrained Loss improves performance at half and quarter resolution but hurts it at full resolution when added to the baseline. That forces a simple check: don't assume full-resolution behavior from gains at lower resolutions. Validate at the resolution you're actually shipping, because the reported trend flips there.
The snippet also doesn't give a dB number. That gap matters. You can cite MatchAttention to justify testing at full resolution, but you can't cite it for the size of any PSNR gain, a flicker-index change, or a block-crawl outcome. When a figure isn't there, the honest move is to keep the mechanism and insist on the measurement instead of pinning a number to the source.
Here's the source-convergence picture, stated once:
| Source | Specific anchor | What it forces |
|---|---|---|
| Emergentmind | 27 May 2026; cross-scale consistency constraints enforce agreement and invariance across spatial, temporal, and frequency domains | Use as a design constraint for multi-scale agreement; don't treat it as a PSNR result |
| MatchAttention (arXiv) | Baseline + Consistency-Constrained Loss improves at half and quarter resolution, degrades at full resolution; no dB value in snippet/td> | Run full-resolution validation; supply the PSNR and flicker measurements locally |
None of the grounded material supplies a video PSNR figure or a flicker figure for a compressed-video sharpening ladder. The Nature item on experiment-guided AlphaFold3 resolves measurement-consistent protein ensembles is not a video-quality measurement source, and the remaining constraint references describe consistency techniques or database rules rather than a CRF ladder result. So the evidence supports the shape of the test, not a precomputed answer.
Read that as a burden-of-proof rule. The sources justify asking whether a pass holds agreement across scales and whether a constrained loss survives at full resolution. They do not justify waiving the local ladder run. Bring your own measured median PSNR, your own static-block flicker index, and your own block-crawl count; then let the gate decide.
Choose the pass by flicker budget
The selection is made by flicker budget, not by PSNR alone. Run the three-clip CRF 23/28/32 ladder, then compare the four passes below; temporal-consistency-constrained sharpening wins only when its median PSNR gain is at least 0.5 dB, its maximum flicker-index increase is at most 10%, and no clip shows block crawl.
| Option | PSNR behavior | Flicker risk | Winner condition |
|---|---|---|---|
| A Per-frame unsharp/RL | can raise PSNR on stills | high on block grid | only if static shot and flicker ≤10% |
| B Temporal-consistency-constrained sharpening | best chance of ≥0.5 dB with constraint | medium-low if gate passes | WINNER when median PSNR ≥0.5 dB and flicker ≤10% |
| C Deblock/denoise only | lower PSNR gain, often <0.5 dB | low | fallback when B fails |
| D Generative restoration | perceptual detail, PSNR unreliable | not specified in this comparison | outside this PSNR/flicker gate |
Option A can raise PSNR on stills, but its per-frame behavior is high-risk on the block grid; it is acceptable only for a static shot that already keeps flicker at or below 10%. Option B is the only pass built to pursue at least 0.5 dB under a temporal constraint; it is the winner when both the PSNR and flicker gates pass. Option C is the fallback: lower PSNR gain, often below 0.5 dB, but low flicker risk. Option D pursues perceptual detail with unreliable PSNR and is not chosen by this gate.
The table is ordered by decision priority, not by PSNR ambition. Flicker risk is the first screen because a sharpening pass that makes the block grid visible has already failed the viewer-facing test, even if its median PSNR is high.
The asymmetry matters: PSNR is judged as a median across the ladder, while flicker is treated as a budget that a single bad clip can spend. A pass that lifts the median but breaks the flicker budget on one clip is rejected; the winner must keep block crawl at zero across all three clips.
If B fails, the decision becomes C rather than a weaker version of A. A remains a special case for static material that already passes the flicker check, and D remains outside the PSNR/flicker gate when perceptual detail is the stated objective.
Use the named source only for the constraint vocabulary: cross-scale consistency constraints are principles that enforce agreement and invariance between multi-scale observations in spatial, temporal, and frequency domains (Emergentmind, updated 27 May 2026). In this comparison, that vocabulary names Option B's constraint; it does not supply a video PSNR or flicker figure.
Costs, limits, and when the gate breaks
The gate has a bounded cost, and it is worth stating that cost plainly before anyone scales it up. Three clips at five seconds each, encoded at CRF 23, 28, and 32, give 3 × 5 × 3 = 45 seconds of test video. Running both the sharpening candidate and the fallback pass doubles that to 90 seconds. At 720p, that is a preflight check, not a full-movie validation run. If your pipeline cannot spare ninety seconds of encode time for a decision that affects temporal behavior, it probably should not be making the decision at all.
The accounting also defines what the gate can and cannot certify. It tells you whether, on those clips and those quantizers, the candidate earns at least 0.5 dB median PSNR while keeping the static-block flicker index increase at or below 10% and block-crawl count at 0/3. It does not tell you that the same result holds for every source type, motion profile, or bitrate ladder. Treat the numbers as acceptance evidence for the ladder you ran, not as a general property of the filter.
Three edge cases are where the rule most often stops being decisive. The table below is the cost-and-limit ledger for those cases: it says when the gate breaks and when the same decision path still produces a valid win. In each row, the fallback remains deblock/denoise unless the stated conditions are met.
| Edge case | When rule breaks | When it still wins |
|---|---|---|
| High motion | Static-block flicker index misses motion flicker, so a clean index reading can hide visible instability. | Motion-masked TCE keeps the flicker index increase at or below 10% and median PSNR gain at or above 0.5 dB. |
| Film grain/noise | Sharpening amplifies grain, so PSNR may drop and the candidate fails on its own terms. | Deblock/denoise first, then rerun the gate on the cleaned source. |
| CRF >32 | The block grid dominates the frame, so sharpening can reinforce compression structure instead of recovering detail. | Only if the measured ladder still clears the same PSNR and flicker conditions; otherwise stay with deblock/denoise. |
The high-motion row is the one most likely to produce a false pass. A static-block metric is built around stationary grid behavior; it is not a complete account of motion flicker. If the clip has fast pans, camera shake, or dense foreground movement, add a motion-masked temporal consistency check before accepting the result. The gate has not been bypassed: the same 10% ceiling and 0.5 dB floor apply, but the measurement has to cover the artifact the clip actually produces.
The grain row is a sequencing problem. Sharpening a noisy source can raise local contrast where there is no stable detail to recover, which can push PSNR down before the flicker question is even reached. Run deblock/denoise first, then measure the sharpening candidate against that cleaned baseline. If the cleaned baseline already meets the display target, the extra pass has to justify itself on the same ladder rather than on an unfiltered comparison.
Above CRF 32, the block grid becomes the dominant visual structure, so the candidate is being asked to sharpen the artifact more than the image. The rule still applies, but the burden of proof gets heavier: the same 0.5 dB median gain, the same 10% maximum flicker-index increase, and 0/3 block-crawl are required. If any one of those fails, the bounded preflight has done its job by rejecting the pass before it reaches a full encode.
Worked gate and if/then rules
Run the gate on three 720p, 5-second clips encoded at CRF 23, CRF 28, and CRF 32. For each clip, compare the sharpened pass against the unsharpened baseline at the same CRF. Record three numbers: PSNR gain in dB, the change in static-block flicker index as a percentage, and whether block crawl is visible. The pass is accepted only if all three checkpoints clear together: median PSNR gain at least 0.5 dB, maximum flicker index increase no more than 10%, and block-crawl count of 0 out of 3 clips. A strong PSNR result does not rescue a flicker or crawl failure.
Copy this worksheet into your notes and fill it from the ladder:
Compute the decision values from the filled rows, not from the best-looking clip. Take the median of the three PSNR gain values, take the maximum of the three flicker deltas, and count the “y” entries in the block-crawl column. Then apply the gate in order: median PSNR gain ≥0.5 dB, max flicker Δ ≤10%, and crawl count = 0/3. If any one of those fails, mark the sharpening pass as rejected for this ladder.
The if/then rules are deliberately strict. If median PSNR gain is at least 0.5 dB, max flicker increase is at most 10%, and block crawl is 0/3, then accept temporal-consistency-constrained sharpening. If median PSNR gain is below 0.5 dB, then reject it even when flicker is flat. If max flicker increase is above 10%, then reject it even when PSNR improves. If any clip shows block crawl, then reject it even when the other two clips are clean. If the worksheet is incomplete, then do not accept the pass; rerun the missing CRF point before deciding.
This is a practical reading of consistency constraints: EmergentMind’s 27 May 2026 overview describes cross-scale consistency constraints as rules that enforce agreement and invariance across spatial, temporal, and frequency domains. In this gate, temporal agreement is tested by flicker and crawl, while quality gain is tested by PSNR across the CRF ladder. The worksheet does not claim that sharpening is generally safe; it only tells you whether this pass earns acceptance under the stated thresholds.
Frequently Asked Questions
What median PSNR gain does the thesis require before accepting sharpening?
The thesis sets the bar at a 0.5 dB median PSNR gain across a CRF 23/28/32 ladder.
Which CRF values make up the ladder used for the PSNR check?
The median PSNR gain is measured across a CRF 23/28/32 ladder.
How many clips should be included in the pre-decision test?
Run the 3-clip CRF ladder before making a call.
What block-crawl result is required for temporal-consistency-constrained sharpening?
Require a block-crawl count of 0 out of 3.
When does temporal-consistency-constrained sharpening lose to deblocking/denoising?
If the peak flicker increase is over 10% or the block-crawl count isn’t 0 out of 3, temporal-consistency-constrained sharpening loses out to deblocking/denoising.
Why can per-frame sharpening turn compressed blocks into flicker?
Compression leaves behind an 8×8 block grid whose edge energy shifts with the quantizer, so a per-frame sharpening kernel treats every block boundary like real detail.
Quick answers
| What median PSNR gain bar does the thesis set for accepting sharpening? | The thesis sets the bar at the 0.5 dB median PSNR gain across a CRF 23/28/32 ladder. |
| What does the reader rule say about the peak flicker index? | The reader rule caps the increase in the peak flicker index at 10%. |
| When does the reader rule allow temporal-consistency-constrained sharpening? | The reader rule says to use temporal-consistency-constrained sharpening only when the block-crawl count is 0 out of 3. |
| What does the reader rule call for before making a call? | The reader rule calls for a 3-clip CRF ladder plus all three acceptance checks. |
| What happens when temporal consistency fails? | If the peak flicker increase is over 10% or the block-crawl count isn’t 0 out of 3, temporal-consistency-constrained sharpening loses out to deblocking/denoising. |
Also worth reading: Upscale blurry compressed video: BasicVSR++ vs cleaning with 78% votes: Upscale blurry compressed video: BasicVSR++ · Fix compressed blurry video: 28dB collapse cleaning vs propagation route: Fix compressed blurry video: 28dB · Transform blurry footage into crystal clear 4K video using artificial intelligence: Transform blurry footage into crystal