Testing 2026 Water VFX Upscalers: VMAF Gate and Flicker Tiebreaker

This convergence proves that the 8× and 128×128→1024×1024 figures are consistently reported, not isolated marketing copy. It does not prove that any output will clear a VMAF threshold of 95.0 or keep the 95th-percentile temporal residual at or below the source baseline on moving-specular water. Qualitative claims of temporal consistency describe an intended property; they are not a substitute for measured frame-to-frame residuals on the exact plate you intend to deliver.

For a reader applying the thesis rule, the evidence supports only a narrow conclusion: the documented operating point is 8× from 128×128 to 1024×1024, and the stated goal includes temporal consistency. Treat every other quality assertion as unverified until you measure it. Run the same 48-frame moving-specular water clip through the candidate at identical output resolution, compute VMAF against your reference, and compute a per-frame temporal residual so you can compare the 95th percentile to the source baseline.

The practical check is binary. If a source does not publish VMAF and flicker scores for your class of footage, it cannot answer the gate or the tiebreaker on your behalf. Accept the converged 8× evidence as a reason to test, not as proof of pass or fail, and let your own matched-resolution measurements decide whether the tool stays in contention.

Three tools, one provisional winner

The documentation cited here supports Adobe VideoGigaGAN as a candidate: the project page and paper describe generative video super-resolution aimed at high-frequency detail with temporal consistency. Topaz, SeedVR2, and Video2X appear only as possible comparison routes in this article; no named source cited here supplies their temporal metrics. Treat them as candidates to test, not ranked results: run the same 48-frame moving-specular water clip through each at identical output resolution, measure VMAF and p95 flicker residual against the source baseline, and keep only routes that meet the gate and baseline before applying the ±0.5 VMAF tiebreak.

Adobe VideoGigaGAN is a documented candidate, not a measured winner. The VideoGigaGAN project page and the accompanying arXiv paper describe it as a generative video super-resolution model designed to produce high-frequency details while maintaining temporal consistency, and they document specific techniques aimed at suppressing temporal artifacts. That gives it a source-backed temporal-consistency claim to test, but the winner must still be the surviving tool with the lowest 95th-percentile flicker residual at matched VMAF within ±0.5.

Topaz and SeedVR2 enter as challengers, not co-winners. No source cited in this article supplies a comparable technical claim about their temporal stability on moving specular water. Their position on the shortlist is therefore conditional: each route must earn placement by meeting the same gate and tiebreaker on the controlled 48-frame test, rather than by reputation or by still-frame sharpness.

This provisional ranking should be treated as a documentation-based starting point, not a final verdict. Run the test with identical input, output resolution, frame rate, and export settings across all three routes so that VMAF and the 95th-percentile temporal residual are directly comparable. Any route that raises flicker above the source plate fails regardless of VMAF, and any route that clears the gate but lacks a documented temporal-consistency claim must still prove lower residual flicker than VideoGigaGAN on the same water clip.

Cost math: 8x means 64x pixels

Linear upscaling multiplies both frame dimensions, so compute and storage costs grow with pixel area, not with the scale factor alone. A 4× upscale produces 4² = 16× the pixels of each source frame, while an 8× upscale produces 8² = 64×. Budget decode time, GPU time, and disk space from output pixels rather than source duration: a longer source at modest resolution can be cheaper to process than a short source pushed to an 8× target.

Apply the same arithmetic to the 48-frame moving-specular water clip used for evaluation. At identical output resolution, every candidate processes the same frame count, but an 8× pass still writes 64 times as many pixels per frame as the source. That multiplier determines whether the test run fits in available VRAM, local storage, and review time before you compare quality metrics.

Treat export logs as the cost record, not the preview window. Before accepting a run, confirm three numbers: output resolution, frames rendered, and wall-clock seconds per frame. Output resolution verifies the intended pixel multiplier, frames rendered confirms the full 48-frame or 240-frame job completed, and per-frame timing exposes the real throughput of the 64× workload.

Reject any tool that cannot show per-frame timing. Preview modes, cached proxies, and partial renders can make a heavy upscale appear inexpensive while deferring the full pixel write. Without a per-frame figure, you cannot extrapolate total runtime for longer shots or compare tools on equal throughput terms.

No grounded price exists for these tools, so do not convert runtime into a currency estimate without verified rates. Instead, calculate total wall-clock time as seconds per frame multiplied by frames rendered, and estimate storage from output resolution, bit depth, codec, and frame count recorded in the export log. Those measured quantities are the defensible cost basis for an 8× workflow in 2026.

Where VMAF misleads on water plates

VMAF is a useful gate for 2026 water-VFX upscaling, but VMAF-only selection breaks on specific water-plate edge cases: dry deck, heavy mist, transparent splash, and compression noise. A model can raise its VMAF score by adding stable-looking detail to still water while simultaneously oscillating on moving spray, producing an average that hides temporal instability. For the 48-frame moving-specular water clip, always pair the VMAF ≥ 95.0 gate with a temporal residual measured on the moving-specular mask; if the p95 flicker residual exceeds the source baseline, the tool fails regardless of VMAF.

These edge cases determine when VMAF can be trusted as the primary selector and when it must be overruled by a targeted temporal check:

On a dry deck plate, the absence of moving speculars reduces the risk that frame-to-frame oscillation drives perceived quality. In this narrow case, a VMAF-only comparison is a reasonable first pass, but it does not replace the baseline flicker check: a tool that introduces new temporal variation into an otherwise stable region still fails the reader rule.

Heavy mist and transparent splash require closer scrutiny because their edges are semi-transparent and constantly evolving. VMAF can miss transparent-edge boil when the artifact occupies a small or low-contrast portion of the frame. A 10-frame side-by-side, stepped through the moving-specular mask, exposes edge crawling or pulsing that a single aggregate score can conceal.

Compression noise presents the opposite trap. An upscaler may denoise aggressively, raising VMAF by producing a cleaner-looking frame while suppressing fine water texture. In that case, verify high-frequency energy retention against the source: if the tool smooths away ripple detail or spray grain, it has traded texture for score and should not pass, even when VMAF and flicker appear acceptable.

Worked water-clip test: 640x360 to 4x

For a worked comparison, use one 48-frame moving-specular water clip and record its source resolution and frame rate before upscaling. Every candidate must render the same 48 frames at the same output resolution with no extra sharpening, so differences come from the upscaler rather than from preprocessing or output size. The 48-frame length is short enough to inspect frame by frame but long enough to include highlight motion; keep the source frame rate unchanged so cadence matches the plate.

Checkpoint 1 establishes the source residual baseline: compute the 95th-percentile frame-to-frame residual on the original 640×360 plate. Checkpoint 2 applies the VMAF gate: the upscaled result must reach at least 95.0 at matched output. Checkpoint 3 applies the flicker gate: the candidate’s p95 residual must be no greater than the source baseline. A candidate that raises flicker above the source plate fails regardless of VMAF. When VMAF scores are within ±0.5, the candidate with the lowest p95 residual wins.

Log the measured source baseline first, then each candidate’s VMAF and p95 residual, then pass or fail against both gates. Those measured values are the audit trail for the tiebreak; if two candidates are within ±0.5 VMAF, their recorded p95 residuals decide the result.

Only candidates passing both gates enter the comparison. If no candidate passes both, the test records no winner for this clip; if several pass, the lowest p95 residual among the VMAF-tied group is selected. This keeps the decision tied to measured flicker rather than to a single sharpness score.

If/then rules for the final tool pick

Apply the canonical rule to the same 48-frame moving-specular water clip rendered at identical output resolution. Compute VMAF and the 95th-percentile temporal residual for each candidate, then keep only candidates with VMAF at least 95.0 and a p95 residual no greater than the source baseline. The final pick is made by the four rules below, not by a single score.

If a candidate clears the VMAF gate but its p95 flicker residual is above the source baseline, reject it regardless of how high its VMAF is. If surviving candidates are within ±0.5 VMAF of the leader—the highest VMAF among candidates that pass both gates—choose the candidate with the lower p95 residual; that lower temporal residual is the tiebreaker.

If every candidate fails the flicker gate at 4×, do not change tools before retesting. Rerun the same clip at 2× and reapply both gates. If only one candidate passes at 2×, use that 2× route for water VFX and reserve the 4× route for static plates.

If the shot has no moving specular highlights, still confirm that the p95 flicker residual does not exceed the source baseline, then rank the VMAF-qualified candidates by VMAF plus high-frequency retention. If the delivery is 4×, keep the 4× route for that shot; otherwise use the highest-VMAF surviving route.

Record the selected rule, the measured VMAF, and the p95 residual in the shot notes so the pick can be audited against the same 48-frame clip later.

Frequently Asked Questions

What VMAF threshold must an upscaled output meet to pass the quality gate?

An upscaled output must clear a VMAF threshold of 95.0 against your reference to pass the gate.

How is the flicker tiebreaker evaluated for moving-specular water footage?

The flicker tiebreaker requires the 95th-percentile temporal residual to be at or below the source baseline on moving-specular water.

What specific test clip should be used when comparing candidate upscalers?

You should run the same 48-frame moving-specular water clip through each candidate at identical output resolution.

Can qualitative claims of temporal consistency replace measured frame-to-frame residuals?

Qualitative claims of temporal consistency are not a substitute for measured frame-to-frame residuals on the exact plate you intend to deliver.

Does the consistently reported 8× operating point guarantee that a tool will pass the quality tests?

The converged 8× evidence should be accepted as a reason to test, not as proof of pass or fail.

Quick answers

Also worth reading: Temporal Consistency Loss: 4K Flicker Reduced 30% (LPIPS/VMAF): Temporal Consistency Loss: 4K Flicker · Video Flicker Fix: 2026 Temporal Consistency Guide with 3 Metrics: Video Flicker Fix: 2026 Temporal · Upscaling Puzzle Pirates 800x600 to 4K: Temporal Consistency Stops 2026 Flicker: Upscaling Puzzle Pirates 800x600 to