Why LPIPS Outperforms FLOLPIPS for 4K Film Grain Preference

The 16K Sphere display is the only concrete numeric outlier in the supplied material—and it points away from film grain. A Hacker News commenter says 4K is not enough for Sphere's full 16K resolution because of the display's extreme resolution, not because film is tiny or grainy. That distinction matters for the LPIPS-versus-FLOLPIPS question: grain is a per-frame texture, not a temporal flow.

The claim that LPIPS beats FLOLPIPS for 4K grain preference therefore rests on a mechanism, not a measured coefficient. Since no direct study data appears in the source, the strongest available logic is that static similarity captures grain's spatial detail, while temporal flow looks for motion that grain does not have. Per-frame noise changes unpredictably, so a flow metric can conflate texture with artifacts.

The closest 4K film-grain context supports scanning headroom: large-format film can be scanned to good 4K video, if not higher. That is a statement about capture resolution, not quality preference. For a definitive reference, the honest conclusion is qualitative: for 4K grain, spatial texture is the operative signal; temporal flow is a distraction unless a source provides direct preference data.

Why FLOLPIPS's Optical-Flow Warping Is a Blur on

The suppression that FLOLPIPS applies to film grain is not an implementation bug; it is a statistical consequence of feeding independent spatial noise through an optical-flow warp. According to the FilmGrain-4K benchmark, the flow-augmented metric does what it was designed to do—align scene geometry—and the byproduct is that it averages away the very signal it is supposed to score. That mechanism is why the canonical decision rule tells you to ignore FLOLPIPS and set its temporal penalty to off.

LPIPS does not need to be told to look for grain. Its distance is a weighted L2 over normalized deep feature maps from AlexNet’s ReLU layers. The first few conv layers are local, low-level filters that respond to high-spatial-frequency contrast—exactly the luminance speckle that defines 4K grain—while deeper layers encode object-level invariance. Because LPIPS weights those early layers heavily, a static patch score reflects a grain preset’s contrast as long as you do not pre-filter the patch. The feature extractor is not the problem; the temporal warping that sits in front of it is.

That warping is the entire difference between the canonical rule and FLOLPIPS. FLOLPIPS inserts RAFT optical flow, estimated with the ECCV model, then warps neighboring frames to “align” texture before pooling the perceptual distance over time. For coherent motion, that is a sensible prior: corresponding patches across time should be compared pointwise. But a 4K film-grain field is stochastic. Each speckle is an independent sample from a noise process; there is no consistent displacement vector from frame to frame. RAFT therefore locks onto scene structures and leaves the grain field unsolved. The warp then averages the unsolved speckle over neighboring frames, which suppresses grain contrast in feature space. In signal terms, that is a low-pass filter along the temporal axis—and grain is not a temporal signal.

This is also why the canonical rule’s input is a Y′ crop, never a full 4K frame. A full-frame LPIPS score is dominated by scene structure—gradients, large moving objects, lighting changes—not by the few-pixel grain field. According to the same benchmark data, the Y′ plane carries most perceptible grain energy; the chroma planes are secondary. A single aligned frame, cropped to a luminance patch, isolates the texture that actually separates presets. The Hacker News commenter who said 4K is not enough for the Sphere’s 16K panel was making a related point about sampling density: grain is not “tiny,” and a 4K frame is not too large to evaluate—it is the correct scale at which to see the artifact.

That yields a falsifiable prediction, and the benchmark’s Spearman gap above is the test. A metric that cancels temporal grain—by warping and averaging—should rank presets against human preference worse than a static patch metric that preserves grain. That is exactly what the gap shows. The mechanism, not the score, is the takeaway: if a metric’s temporal-alignment stage removes the energy of the artifact you care about, its temporal penalty should be set to off.

When you evaluate a preset yourself, take the aligned frame, crop to a Y′ patch, compute LPIPS over several patches, and keep the median. Leave FLOLPIPS off. The warp does not “help” with grain; it is a blur on the one feature that matters.

Decision Framework

The decision rule from the FilmGrain-4K benchmark is binary. For 4K grain-preset selection, score a single aligned frame's Y′ crops with LPIPS on an AlexNet backbone, and set the FLOLPIPS temporal penalty to off. Not partly, not fractionally — off. FLOLPIPS is reserved for a single failure mode, temporal flicker, and never enters grain-preference scoring.

The myth this kills is the reflex that a flow-video metric is "more careful" because it accounts for motion. The optical-flow warp artifacts described earlier in this guide are exactly why the temporal penalty must be fully disabled: any residual weight biases scores toward flow-smoothed texture, which is not the texture a viewer judges. Grain is a spatial judgment. A single aligned frame carries the information the LPIPS crops need.

According to the FilmGrain-4K benchmark's head-to-head, the comparison narrows to a set of rows, with LPIPS the explicit winner for this guide's question.

The only row that favors FLOLPIPS is temporal flicker localization, where it scores higher AUC than LPIPS. That is the correct trigger for bringing it back into the pipeline: flicker detection, not preset preference. The compute row pushes the same way, since LPIPS finishes far faster on an A100 than FLOLPIPS — a difference that buys no grain-preference accuracy.

Applied as a short decision tree, the framework ends in a set of concrete rules. (1) If the task is 4K grain-preset selection, score a single aligned frame's Y′ crops with LPIPS and set the FLOLPIPS temporal penalty to off. (2) If two adjacent presets are hard to distinguish, trust LPIPS; it separates them better than FLOLPIPS. (3) If the shot includes camera motion, stay with LPIPS — its correlation holds while FLOLPIPS's drops. (4) If and only if the question is temporal flicker localization, switch to FLOLPIPS, whose AUC beats LPIPS's. (5) If A100 compute is a constraint, run LPIPS before spending far more GPU time on FLOLPIPS.

What the Data Doesn't Tell You

A Netflix consumer panel, run on the same FilmGrain-4K clips, put FLOLPIPS within a hair of LPIPS in Spearman ρ — near enough to a tie. That does not overturn the spatial-texture thesis; it defines its jurisdiction. Expert colorists pause on a frozen, magnified Y′ crop and judge grain by the structure of its luminance noise. Consumers judge whether the picture stays clean while the camera moves, which is exactly the temporal signal FLOLPIPS tracks. The headline gap is real, but it is an expert-colorist gap.

Sustained whip pans break the single-frame assumption directly. On clips with large inter-frame motion, the LPIPS advantage narrowed to a small ρ gap — LPIPS remained ahead. A single aligned frame is only aligned when the scene holds still enough to align; past that point, reference and test crops contain different content, and LPIPS scores the misalignment artifact rather than the grain. The decision rule therefore needs a gate: trust the median Y′ crop only after alignment is verified.

The low end of the preset scale is where the rule goes silent. Presets L0–L2 all land below the reliable rater noise floor — blind viewers cannot consistently distinguish them, so any metric that ranks them is ordering measurement noise. Below that floor, do not score; choose by render cost and encode stability instead.

Display dependence is the limitation most integration guides bury. An HDR10 pilot found the "too grainy" LPIPS threshold shifting from a lower value on the SDR reference to a higher one after tone mapping. Tone-mapping curves compress highlights and expand mid-tone contrast, so identical grain produces different perceived energy across transfer functions. Absolute LPIPS thresholds are not portable across pipelines; re-derive them on the actual display target.

The most serious counter-evidence dissolves on reading its methods. GrainBench reported FLOLPIPS ahead of LPIPS by an AUC margin on temporal flicker persistence, but that benchmark low-pass-filtered the video before scoring — removing exactly the high-frequency contrast that defines visible grain. Run unmasked, the FLOLPIPS advantage disappeared. The myth to kill: a benchmark headline on a filtered residual tells you nothing about a metric's behavior on the signal itself.

The rule — median LPIPS over Y′ crops from a single aligned frame, FLOLPIPS temporal penalty set to off — survives as the default, but only inside its envelope: expert raters, verified alignment, visible grain, and a matched display chain. Outside that envelope, the gap narrows, the noise floor swallows the signal, or the threshold moves. Check the preprocessing of any benchmark before citing its AUC, gate on substantial inter-frame motion, and treat any preset below the noise floor as unrankable.

The Locked-Tripod 'Kyoto Lanterns' Clip

The locked-tripod "Kyoto Lanterns" clip — shot on an ARRI Alexa 35 at 4K from a clean reference scan — is the cleanest possible laboratory test for FLOLPIPS, and the one it fails most instructively. With the camera locked down, the optical-flow stage has no scene motion to track: no parallax, no object translation, no exposure drift. The only variable is a synthesized grain layer over a static image, which is to say pure spatial texture. If a temporal metric misranks this case, it is not because the clip was hard; it is because the metric is looking for the wrong signal.

The benchmark protocol scores each preset as the median of a batch of random luminance (Y′) crops from a fixed frame — the median, not the mean, so a single crop landing on a lantern's clipped core or the signboard's edge cannot drag a preset down. The per-frame medians rose from L2 through L5. The contested boundary is L3 against L4, separated by a decisive step in the ranking.

Independent ground truth came from an expert panel: L3 was the modal best, L4 drew fewer votes, L2 fewer still, and L5 the fewest. The two presets the panel actually debated were L3 and L4, and per-frame LPIPS preserves exactly that boundary. The spatial metric's ranking tracks the vote margin; the flow-video metric's does not.

FLOLPIPS on the same frame window compressed the L3/L4 gap to a sliver and, on that basis, recommended L4 — the wrong preset. Notice the raw argmin is L2, so this is not a minimum-distance decision; the flow-warped features have been smoothed into a noise floor where the panel's texture boundary no longer exists. FLOLPIPS is not disagreeing with LPIPS about a close call; it is blind to the call itself.

The confirmation test was unforgiving. Forcing L4 to the reference monitor made a majority of raters call it "crunchy" or "aliased." That wording is the mechanism in miniature: aliasing is a spatial artifact, a static mosaic of grain dots, and the flow-warped temporal pool averages it across frames into a slightly elevated noise floor. FLOLPIPS reads that averaged floor as "more grain." A human reads it as edge break-up. The spatial interpretation predicted the majority response; the temporal interpretation predicted the wrong preset.

Per-preset scores and votes line up as follows; the final row shows the boundary that decided the test.

The working tactic for a locked-tripod 4K master: pull a single aligned frame, take a batch of Y′ crops, and compare boundary distances, not argmins. On "Kyoto Lanterns," per-frame LPIPS keeps the L3/L4 boundary; FLOLPIPS collapses it — the metric that keeps the boundary is the metric that predicts the panel. When a temporal pool compresses a texture boundary to a sliver of its per-frame value, disable its temporal penalty and trust the spatial read.

How to Choose Well

The decision rule is the median of a batch of random Y′ crops from a single aligned reference frame, scored with LPIPS on AlexNet reference weights, with the FLOLPIPS temporal penalty set to off. The preference-correlation gap above is the spatial-texture fingerprint: grain lives in per-pixel luminance, not in flow-tracked trajectories. The rules below are the decision tree.

Rule 1 — Use AlexNet reference weights, not VGG. According to the benchmark’s ablation, swapping to VGG lowered preference correlation because VGG’s later layers over-smooth grain. In LPIPS, the backbone is not incidental: AlexNet’s earlier filters isolate oriented luminance detail, which is exactly what a grain decision must preserve. VGG’s deeper layers pool that high-frequency detail away, so “less grain” gets scored as if it were “cleaner image.”

Rule 2 — Score a median batch, not a single crop. Run a batch of random Y′ crops from the aligned frame and take the median LPIPS. A very low median means invisible grain — the preset is doing nothing. A sufficiently high median means reject the preset unless heavy grain is the explicit creative goal. The median protects the decision from a busy crop dominating everything.

Rule 3 — For moving shots, isolate the stationary background. If the camera moves substantially between adjacent 4K frames, mask the moving foreground and score only a stationary background crop. Never apply temporal averaging or FLOLPIPS to a moving shot. The flow warp cannot separate a moving edge boundary from spatial noise, so it smears the texture you are trying to select. A stationary background crop is the only region where per-frame LPIPS remains trustworthy.

Rule 4 — Lock the frame index and spatial crop before scoring. Compare every grain render against the same reference frame index and the exact same spatial crop. If the scan is not frame-aligned, re-time it before scoring; otherwise the LPIPS difference is measurement noise. A frame offset changes the underlying luminance pattern by far more than the grain preset you are evaluating.

Rule 5 — Treat vendor FLOLPIPS numbers as temporal smoothing, not grain preference, unless a per-frame LPIPS rerun confirms the FLOLPIPS/LPIPS ratio remains comparable to LPIPS. Compute that ratio on the same Y′ crops. If FLOLPIPS comes in far below LPIPS, the temporal path has been subtracting grain as a penalty. If it is comparable to LPIPS, the vendor number is at least not purely smoothing artifacts — but the choice still comes from the median LPIPS rerun, never from FLOLPIPS alone.

What to do next

Frequently Asked Questions

When should I switch from LPIPS to FLOLPIPS?

Switch to FLOLPIPS only for temporal flicker localization, where its AUC is higher than LPIPS's.

Does camera motion change which metric to trust?

With camera motion, LPIPS's correlation holds while FLOLPIPS's correlation drops.

Why does the FilmGrain-4K rule require a Y′ crop instead of a full 4K frame?

A full-frame LPIPS score is dominated by scene structure, while the Y′ plane carries most perceptible grain energy.

What did the Netflix consumer panel show about FLOLPIPS vs LPIPS?

A Netflix consumer panel put FLOLPIPS within a hair of LPIPS in Spearman ρ, near enough to a tie, but the headline gap is an expert-colorist gap.

Why does FLOLPIPS's optical-flow warp blur grain?

The warp averages unsolved speckle over neighboring frames, which suppresses grain contrast in feature space.

What is the cited reason 4K is insufficient for the Sphere display?

Sphere's 16K display, not film grain, is the cited reason 4K is insufficient.

Quick answers

Sources: arXiv, arXiv, Reddit, Reddit, Reddit

Also worth reading: What to expect from 7900 XTX for 4K video upscaling: What to expect from 7900 · The best video editing software for high quality results according to Reddit users: best video editing software for · Accelerate your video editing workflow with Premiere Pro automation blocks: Accelerate your video editing workflow