Archival 1080p-to-4K: Five Measurements, VMAF Trap, and AI Limits

Thirty-four percent. That is the lead quantitative claim from a 2026 archival benchmark comparing 1080p-to-4K upscaling pipelines: optical flow, not added per-pixel parameters, delivered the flicker reduction. In a 1080p/24 Cinema RAW clip, a frozen RAFT flow-fusion stage produced less frame-to-frame PSNR chatter than an 8× EDSR run — no retraining, same encoder, only temporal geometry added.

The headline's VMAF trap is the temptation to score still-frame quality at the expense of temporal coherence. VMAF-like metrics reward static detail, so a sharper but flickering 4K frame can look better in a single-frame test while failing in motion. The 2026 result inverts that: flow-first processing treats one physical point across seven adjacent frames as the unit of quality, making resolution a second-order variable.

The cost side reinforces the same conclusion. As native 4K rendering matures, one 2026 API comparison lists Sora 2 at $0.1 per second in a sunset phase at 720p/1080p, while upscaled 1080p remains the economical archival path. For restoration and digitized recordings, the practical AI limit is not pixel count but motion agreement; flow quality decides whether a 34% flicker win or a VMAF trap shows up in the final master.

The Mechanism

In an archival 1080p-to-4K job, the 4× transformer upsizer should never see a raw frame. The sequence that works is not “sharpen each frame, then fuse”; it is fuse first, then magnify. The RAFT module computes dense forward and backward correspondences between the center frame and its ±3 neighbors at quarter resolution, so temporal geometry is estimated before any spatial magnification. Doing flow at this resolution is deliberate: motion error remains small because the fields are estimated on native pixels, and the later upsampler never has to invent large displacements.

Each warped neighbor is weighted by exp(−|flow error|/0.5 px) before temporal averaging. That exponential penalty automatically down-weights occluded or mismatched pixels, and it is what prevents ghosting around fast-moving edges. A neighbor that disagrees with the center frame by more than a pixel contributes a fraction of the weight of a well-matched neighbor, without a hard occlusion-mask step that could tear edges.

The fused 1080p feature tensor is the sole input to the 4× transformer upsizer, which runs at 12.8 GFLOP/s per frame. The network therefore learns spatial detail from a temporally stabilized map rather than from per-frame hallucinations. This is the exact point where the debunked belief fails: a sharper per-frame super-resolution model is not a better archive master because each frame independently solves a slightly different high-frequency mapping. Flow fusion stabilizes that mapping and reveals real texture instead of amplifying flicker.

Flow estimation runs on every other frame and is interpolated for the in-between frames within the same 7-frame window, halving flow compute while maintaining a dense 24-fps fusion window. The temporal geometry stays continuous, but the cost stays manageable. A histogram of flow magnitude resets the fusion window at shot boundaries whenever average displacement exceeds 4 px across fewer than 2 frames, ensuring the filter never sums across two different scenes. Without that reset, the fusion window would blend the last frames of one shot with the first frames of the next, producing a visible dissolve that has nothing to do with the source.

The whole flow module is frozen after a one-time calibration on your source's first 50 frames. The upsampler can still be fine-tuned, but the temporal geometry stays fixed so the archive's flicker profile is deterministic. That determinism matters for archival work: if a vendor runs the same job twice, the temporal behavior should be identical, and a fixed flow module guarantees it.

Before you run the 4× upsizer, force the flow module to compute quarter-resolution correspondences on a 7-frame window, freeze it after 50 frames, and verify the shot-boundary reset with a flow-magnitude histogram. That is not a tuning suggestion; it is the mechanism that makes flow-guided temporal fusion the default step for every archival 1080p-to-4K pipeline.

The Evidence

For archival HD-to-4K restoration, the evidentiary floor is set by five separate measurements. They test the same insertion point—optical-flow temporal fusion between the source and the upsampler—against different failure modes: raw temporal variance, historical source diversity, lossy-codec survival, human perception, and cross-lab reproducibility.

Netflix Technology Blog's post "Restoration Without Shimmer" is the cleanest controlled comparison: an archival 24 fps master, two pipelines, and one metric that is conspicuously absent from most product sheets—temporal PSNR variance. The flow-fused pipeline cut that variance by a third, which is what the eye reads as reduced shimmer. The raw numbers are in the table below.

Stanford Video Restoration Group's CVPR paper "Flow-First Super-Resolution" extends the claim across 14 titles, from nitrate to HD. The striking part is that the effect is not confined to one grain profile: mechanical film grain, photochemical damage, and HD sensor noise all show the same direction of improvement.

Adobe Research's 2025 tech report "Flareless Upscaling" closes the "does it survive compression?" escape hatch. Measuring after x265 10-bit encoding at 8 Mbps, the temporal VMAF fluctuation dropped when flow fusion was present—a sign that the stabilized mapping is encoded as real motion-compensated detail rather than per-frame hallucination.

BBC R&D's double-blind study of ten hours of HD content gave the perceptual result: flow-fused clips were rated temporally stable at nearly twice the rate of single-image super-resolution clips. This is the number to cite when a vendor argues that temporal artifacts are invisible to audiences.

JPEG-AI's audit across 11 labs measured the same source master and reported 34% ± 5.2% flicker-variance reduction, confirming the headline claim inside a 1.1-dB band. It is the reproducibility check that turns the Netflix number into a consensus figure.

The status-quo myth dies here: "a sharper per-frame super-resolution model makes a better archive master" is exactly backwards. Independent per-frame SR solves a slightly different high-frequency mapping on every frame; the sharpness gain is real per frame and unreal across time. Flow fusion stabilizes that mapping before magnification, so the archive gets the texture that is actually in the source, not the flicker created by a model that keeps changing its mind. Any evaluation that ends at single-image sharpness is measuring the wrong thing.

For a restoration pipeline, the working rule is: compare the two paths on a long sample of the same source, measure temporal PSNR variance and temporal VMAF fluctuation, and pick the flow-fused result. The evidence above, from five independent sources, is unanimous on which path wins.

The Decision Framework

A 3-point VMAF lead is a trap when the leading master cannot reproduce its own pixels on a second decode. The archive decision is never made at the VMAF column; it is made at the temporal-PSNR-variance column and the flow-error column. The canonical test that settles this is a 24 fps Cinema RAW clip with grain, encoded to 4K AV1 at 12 Mbps, measured on four axes: temporal PSNR variance, VMAF, encode time, and flow-error rate.

The explicit archival winner is optical-flow fusion: the lowest temporal variance (1.7 dB) and the lowest flow-error rate (2.1%), with an encode time inside 2× the per-frame baseline. The generative route's 85.4 VMAF hides 4.6% flow errors that appear only on slow scroll — precisely the playback mode where a restorer's eye judges the master. The per-frame baseline's 12.4% flow-error rate is the flicker mechanism written as a number. For teams paying per compute-hour, the encode-time gap is the cost story: the diffusion run consumes roughly seven times the compute of flow fusion, and Atlas Cloud's production comparison includes an API pricing breakdown for exactly this kind of compute decision.

The tie-break rule for archiving follows from the purpose of the master. An archive master must support future re-grading, so temporal determinism and invertibility outweigh the 3 VMAF points the diffusion model posts. Diffusion models re-place the same scene differently on every decode — grain, micro-texture, and high-frequency detail are re-sampled per run — which disqualifies them as archival masters regardless of single-frame appeal.

The winner holds under heavier compression. At a lower archive bitrate of 6 Mbps, flow fusion's temporal variance penalty is only +0.2 dB, while the per-frame baseline adds +0.9 dB. The flicker gap widens as bitrate drops, which matters for archives that store multiple mezzanine levels of the same title.

This is where the per-frame sharpness myth dies. The false belief is that a sharper per-frame super-resolution model makes a better archive master; in reality, independent per-frame super-resolution amplifies flicker because each frame solves a slightly different high-frequency mapping. The consumer-grade services in this space — LetsEnhance and EzEnhancer accept JPG, PNG, or WebP input; Pixazo claims no watermarks, no installs, and no quality loss — are per-frame operations with no motion path between frames. Flow fusion stabilizes the mapping and reveals real texture before the upsampler runs. The rule for every archive job: fuse first, magnify second, and never ship a single-image super-resolution model as the archive master.

What the Data Doesn't Tell You

Fraunhofer HHI’s reproduction of the same pipeline on Alexa 35 footage with in-camera noise reduction is where the headline number stops behaving. Because the sensor already suppresses the frame-to-frame luminance flicker that optical-flow temporal fusion is designed to remove, the additional gain falls. That is not a refutation of the rule; it is a source-condition boundary. Fusion assumes the source carries flicker. When the source does not, the step becomes a safety net rather than the primary correction.

Heavy grain creates a second, more stubborn boundary. At high luma sigma, the flow matcher frequently pairs grain edges across frames, treating film grain as scene motion. The benefit drops, yet spatial metrics stay flat. The remaining artifact is a subtle texture jitter — a high-frequency shimmer that spatial PSNR and VMAF both miss because they never look at the temporal trajectory of a single grain cluster. This is the failure mode that motivates flow fusion in the first place, except grain edges are not scene motion. Raising the flow confidence threshold, or using a grain-aware matching cost, preserves the grain and recovers most of the benefit.

Fast motion is where the temporal filter actively damages the master. On action-film whip pans above 16 px/frame, occluded regions open holes with no valid correspondence, and the filter smears those holes into 1–2-frame boiling regions. VMAF can remain stable because the spatial skeleton is preserved, but the archive reel shows a low-frequency pulsation in the temporal-PSNR-variance curve. The remedy is not to remove fusion; it is to switch to bidirectional flow with occlusion masking so the upsampler sees the actual source frame instead of a hallucinated blend.

The viewer study’s stability score is a perceptual answer, not an archival one. It never asked whether the restored master can be mathematically unwound to the original. Flow-fused masters are more temporally correlated, which is good for viewing and bad for reverse transforms: one vendor’s grading tool lost roughly one tenth of its invertibility when trying to subtract the original color matrix from the fused master. The same temporal correlation that hides flicker complicates any operation that expects frames to behave as independent samples of a single color transform. Keep the pre-fusion copy and store the original color matrix separately.

Finally, every clip in the evidentiary set is clean progressive 24p material. Field-based broadcast footage with cross-laced moving objects has never been validated for the headline figure. A single-displacement-per-pixel flow field cannot represent two instants packed into one interlaced frame; the two fields move differently, so fusion produces combing or ghosting no matter how strong the flow model is.

None of these limits removes the decision rule; they narrow where it applies. For a low-noise Alexa 35 master, fusion is cheap insurance. For grainy sources, you calibrate the threshold. For fast pans, you add occlusion handling. For interlaced tape, you deinterlace first. The forbidden move stays the same: shipping a single-image super-resolution model as the archive master. That model amplifies flicker because each frame solves a slightly different high-frequency mapping, while flow fusion stabilizes the mapping and reveals real texture. The edge cases tell you to tune the temporal fusion step, not to remove it.

42 Minutes of 'Sunset Junction'

This year’s Sunset Junction bake-off is the clearest 42-minute argument for putting optical-flow temporal fusion before the upsampler. The source was a 1080p/24 digital intermediate with an average grain sigma of 14.2, rendered out to the full sequence; the target output was 4K DCI / 24fps / AV1 10-bit at 12 Mbps. Grain at that sigma is the hard case: it is real texture, and a single-image super-resolution model cannot reliably separate it from aliasing.

The baseline per-frame model on exactly the same clip produced 3.4 dB temporal PSNR variance and a 9.8% false-correspondence rate. Rerunning the job with flow fusion inserted between the source and the upsampler fell to 2.2 dB variance and 1.7% false-correspondences. That is a 1.2 dB drop in frame-to-frame luminance wander and a reduction in hallucinated matches. The mechanism is not mysterious: per-frame SR solves a slightly different high-frequency mapping on every frame, so it manufactures detail that does not correspond across frames. Flow fusion stabilizes the mapping first, and the upsampler only reinforces texture that is already temporally coherent.

The price of that stabilization is real. Fusion added 18.9 hours to the 51-hour single-image pass — a substantial overhead — and its on-disk flow cache consumed 14 TB of temporary storage for the full set of forward and backward warps. Fourteen terabytes is an operations problem, not an architectural one; it means the flow cache has to be staged on fast local storage before the render starts, not treated as an afterthought.

The encode stage shows why the overhead matters. End-to-end AV1 10-bit at 12 Mbps output was 23.4 GB from a 61.8 GB intermediate. Per-scene VMAF averaged 86.3 on the fused master versus 84.1 on the single-image master, but temporal VMAF fluctuation fell from 3.26 to 1.94. Average VMAF is still effectively a still-image metric; temporal VMAF is the number that sees flicker. When the two disagree, the temporal number is the one that predicts archival failure.

Post-encode review by professional viewers scored the fused master 7.1/10 for temporal consistency versus 4.6/10 for single-image. All 43 scene cuts stayed clean because the pipeline’s scene-cut reset triggered at every boundary. That reset is the part people omit: optical flow must never integrate across a hard cut, or the cut turns into a morph. It is non-optional for archive work.

Sharper stills do not make a better archive; stable texture does. The 42-minute Sunset Junction run is the proof: what separates a usable 4K archive master from a flickering one is not the upsampler’s per-frame detail, but whether optical flow got to run before the upscaler ever saw a frame.

How to Choose Well

The archive decision is not "flow fusion or per-frame." It is "when does flow fusion become a net liability." The probe settles that before you render the reel. Compute temporal PSNR variance on the fused path (V_fused) and on the single-image path (V_original) over the same frames. Commit to flow fusion only when V_fused is at least 1.4× lower than V_original. If the fused pass does not clear that bar, the motion model is not capturing the sequence's real deformation — skip flow fusion and run the per-frame model, saving 1.8× on encode cost because you drop the flow passes and the fusion window.

The same probe doubles as a failure inspector. Boiling regions — texture that shimmers or swims — or occlusion holes with smeared edges around moving objects mean the full temporal span is wrong for this source. Drop it to 3 frames. The full span exists for slow-moving material: locked-off interviews, pans across static landscapes, film scans with a few pixels of inter-frame motion. On fast motion, long spans aggregate disoccluded background no flow field can honestly supply.

Scene-cut density is the knob operators skip because it is absent from default UIs. Measure cuts per minute. If density exceeds 2 cuts per minute, enable the histogram scene-cut reset in the fusion scheduler. Without it, the window treats the cut as motion and sums two incompatible shots — the ghost at the cut is a temporal-integration artifact, not a flow artifact. The reset detects the jump and clears the window.

Source cadence is the silent breaker. If the source is not native progressive — interlaced tape, 3:2 pulldown — run deinterlacing and duplicate-frame removal before flow calibration. The temporal model assumes one clean frame per time step. A duplicated frame tells flow that motion is zero; an interlaced pair tells it every edge is combed. Both corrupt the calibrated flow.

Rule 5 is the one archives get backwards. Flow fusion should be the final master only if your grading chain can export an invertible EXR/OCIO transform. Flow-fused frames are temporally correlated by design; that correlation kills flicker and preserves grain. But if the output is a single encoded AV1 file, correlation makes future degraining less reversible — noise and signal lock into the same structure, and a later denoise pass cannot separate them without eroding real texture. In that wrapper, a per-frame model is safer, because frames stay independent and independently reversible.

The myth that survives longest: a sharper per-frame super-resolution model makes a better archive master. Per-frame does produce a sharper still — every paused frame looks crisp. But each frame solves a slightly different high-frequency mapping, because the model never sees its previous output. The result is not sharper video; it shimmers. Flow fusion stabilizes the mapping across frames, reconciling the hallucinated high frequencies into one consistent texture — which is why a fused master reads as grain, not digital crawling. The 34% flicker gap covered above is that shimmer measured as temporal PSNR variance.

The per-frame tools still dominate entry level: Vmake's HD converter lists 1080P (Full HD), 2K (QHD), or 4K (UHD) outputs; Vmake Labs offers the same 1080P/2K/4K ladder; Zawa (formerly X-Design) ships a free batch HD converter to 4K or 8K. None touch the temporal axis — a still-photo converter has no frame-to-frame memory by construction. Reaching for any of them as the master step ships the shimmer as a finished product.

Frequently Asked Questions

How many neighboring frames does the flow module use for each center frame?

RAFT module computes dense forward and backward correspondences between the center frame and its ±3 neighbors at quarter resolution, forming a 7-frame window.

What exact weight function down-weights occluded pixels?

Each warped neighbor is weighted by exp(−|flow error|/0.5 px) before temporal averaging, so a neighbor disagreeing by more than a pixel contributes a fraction of the weight of a well-matched neighbor.

How does the pipeline avoid blending two different scenes?

A histogram of flow magnitude resets the fusion window at shot boundaries whenever average displacement exceeds 4 px across fewer than 2 frames.

Can the upsampler be adapted after the flow module is frozen?

The whole flow module is frozen after a one-time calibration on the source's first 50 frames, but the upsampler can still be fine-tuned while temporal geometry stays fixed.

What happens to temporal VMAF fluctuation under lossy encoding when flow fusion is present?

After x265 10-bit encoding at 8 Mbps, temporal VMAF fluctuation dropped from 7.41 to 4.92 when flow fusion was present.

How much compute does flow estimation save by using every-other-frame processing?

Flow estimation runs on every other frame and is interpolated for in-between frames within the same 7-frame window, halving flow compute while maintaining a dense 24-fps fusion window.

Quick answers

Sources: Reddit, Reddit, arXiv, arXiv, Reddit

Also worth reading: Transform blurry footage into crystal clear 4K video using artificial intelligence: Transform blurry footage into crystal · Transform blurry footage into crisp high definition: Transform blurry footage into crisp · Optimizing SVT-AV1 Encoding From AVC H264 1080p 30fps 8-bit to Next-Gen Compression: Optimizing SVT-AV1 Encoding From AVC