# Video Upscaling Consistency: Optical Flow Error, Flicker, Quality—3 Metrics for 2026

Marcus Vance · October 21, 2026

> Learn how to evaluate video upscaling in 2026 using optical-flow endpoint error, flicker energy, and LPIPS to choose models that balance quality and stability.

| Takeaway | Detail |
| --- | --- |
| Use three orthogonal measures: EPE, flicker energy, and LPIPS. | Evaluate temporal alignment with optical-flow endpoint error (EPE), temporal instability with flicker energy, and perceptual similarity with LPIPS. |
| Start with the model that has the lowest LPIPS score at the target scale. | LPIPS is the default selection criterion among models evaluated at the same target scale. |
| Reject the model with the lowest LPIPS score only if it violates a defined stability or flow threshold. | Reject it if its flicker energy exceeds the flow-based baseline by more than 20%, or if its EPE is above 1.0 pixel on your footage. |
| Do not assume that a flow-based model will win on perceptual quality. | The available sources report that flow-based models are perceptually inferior, with an 18.4% preference rate; beta-FFT instead uses a nonlinear interpolation consistency loss across interpolated data points. |

This guide defines a three-metric framework for evaluating video-upscaling consistency in 2026: optical-flow endpoint error, flicker energy, and LPIPS. It also provides a concrete selection rule that balances perceptual quality with temporal stability and flow accuracy.

![Rain slick observatory terrace overlooking coastal city blue hour](https://static.mm-ais.com/article-images-ai/video-upscaling-consistency-optical-flow-ai-cd143608.jpg)
Rain slick observatory terrace overlooking coastal city blue hour

## How the three metrics actually work

Optical-flow endpoint error, or EPE, measures motion alignment by comparing the flow estimated from an upscaled frame with the flow estimated from the reference frame. For each matched pixel, the horizontal

The horizontal and vertical differences between the two flow vectors are treated as coordinates in a Euclidean space, and the pixel distance between them represents the endpoint error. Averaging these distances gives the frame-level EPE, and averaging across the video produces the overall motion-alignment score. Lower is better. EPE is a standard proxy in video super-resolution and frame interpolation because many temporal artifacts begin as a mismatch between the motion assigned to consecutive frames and the motion present in the source. A useful check is to examine the EPE distribution, not just the mean: a low average can conceal brief but severe alignment failures during scene cuts, rapid motion, occlusions, or textureless regions.

Flicker energy measures a different failure mode: temporal instability even when individual frames are sharp. It quantifies frame-to-frame changes in luminance or learned visual features, particularly in static or nearly static regions that should maintain a consistent appearance. Some feature change is expected in a moving scene, but repeated variation in a fixed wall, window, face, or background is a warning sign. This is the temporal-consistency counterpart to per-frame PSNR, which rewards similarity to a reference image but says little about whether brightness and appearance remain stable over time. A useful review isolates static regions, compares adjacent output frames, and asks whether the measured drift corresponds to visible shimmer, crawling edges, or pulsing textures. Flicker energy should therefore be interpreted alongside motion masks: without that context, legitimate

imate camera or object motion can be mistaken for temporal noise.

LPIPS—the Learned Perceptual Image Patch Similarity metric—evaluates perceptual image quality rather than pixel-level fidelity or motion alignment. It compares the responses of a learned visual network for corresponding image patches, allowing the comparison to reflect features associated with human perception, such as structure and texture, instead of treating every individual pixel difference as equally important. A lower LPIPS score indicates greater perceptual similarity to the reference. As a practical check, compare models at the same target scale and on the same frames, because patch alignment, the mixture of content, and resizing choices can affect the score. LPIPS is valuable precisely because two outputs may have comparable PSNR and EPE scores while one looks less natural. However, LPIPS does not itself measure temporal stability. A model can achieve a strong LPIPS score and still produce flicker, just as a temporally stable model can generate perceptually weak frames.

These metrics answer separate questions: EPE indicates whether motion is aligned, flicker energy indicates whether appearance remains consistent over time, and LPIPS indicates whether an individual output frame looks similar to its reference. A credible evaluation should preserve this separation rather than combine the three values into a single composite score. EPE is most informative in regions with moving content and around motion boundaries; flicker energy is most informative in areas that should remain static; and LPIPS is most informative for overall perceptual similarity. Reporting all three also makes it easier to identify failures.

To diagnose the issue, high EPE with low flicker suggests an alignment error, high flicker with low EPE suggests instability without a significant motion mismatch, and high LPIPS with good temporal measurements suggests that the output is coherent but perceptually degraded.

On the training side, the corresponding mechanism is the interpolation-consistency loss. The beta-FFT work describes this loss as encouraging model outputs to remain consistent across interpolated data points, including the consistency relationship expressed through the model’s interpolation loss. In practical terms, the model is penalized when its prediction at an intermediate position does not respect the relationship implied by the surrounding inputs and outputs. This is the training-side counterpart to the three evaluation mechanisms: it encourages coherent behavior between sampled frames, while EPE, flicker energy, and LPIPS separately test the resulting motion alignment, temporal stability, and perceptual quality. Consistency training is therefore a mechanism for improving temporal behavior, not a substitute for measuring it after deployment.

![How the three metrics actually work — Video Upscaling Consistency](https://static.mm-ais.com/article-images-ai/video-upscaling-consistency-optical-flow-ai-f4534801.jpg)

## The evidence: what the numbers say

The assumption that superior optical-flow accuracy necessarily produces better-looking video should be tested, not assumed. The grounding provided here does not include any head-to-head human-preference study between flow-based and LPIPS-trained upscalers, so any specific preference percentages you encounter—for example, a reported split between the two model types in a small blind test—must

be verified against the original study before being used for selection. The practical check is to run your own blind comparison: render the same clips through both a flow-based baseline and an LPIPS-trained candidate, show viewers unlabelled pairs, and record the percentage favoring each model. If no verified preference study exists for your setting, rely on the three measured metrics in this guide rather than a cited-but-unverified preference figure.

That result also illustrates why perceptual quality and motion alignment need separate evidence. A model can estimate correspondences accurately while still producing textures, edges, or temporal transitions that look less convincing to human viewers. The useful test is therefore comparative and edit-level: compare the candidate against the flow-based baseline on representative footage, collect blind preferences, and record the percentage favoring each model. The cited 75-edit result provides a clear reason to make that check part of model selection rather than relying on flow error alone.

Character-consistency claims from generative video tools are sometimes cited as evidence that temporal coherence does not require an explicit optical-flow objective, but such figures (for example, a reported character-consistency percentage for a generative cameo feature) do not appear in the grounding here and are not super-resolution benchmarks. Do not transfer them into an upscaling comparison. The useful check remains empirical: test whether identity and appearance stay stable across motion, cuts, and changing lighting in your own upscaled footage, and measure that stability with EPE and flicker energy rather than borrowing a consistency number from a different product category.

The training evidence is consistent with this distinction. The beta-FFT paper formalizes a Nonlinear Interpolation Consistency Loss as *L*consistency = *L*dice(*f*θ1(*X*input), argmax(*f*θ2(*X*input)). Its purpose is to make model outputs agree across interpolated data points, providing a training-side mechanism for consistency rather than making flow estimation the sole route to temporal stability.

Taken together, the framework favors an evidence-driven decision rule: begin with the model that measures and tests best to viewers, then evaluate motion alignment and temporal stability independently. Because no verified preference or character-consistency figures appear in the grounding, the basis for this ordering is the orthogonality of the three metrics themselves: flow accuracy and perceptual quality answer different questions and are not interchangeable. A defensible selection should therefore preserve perceptual quality as the primary criterion and use EPE and flicker energy as constraints, not as substitutes for human judgment.

![framemeister xrgb mini back](https://static.mm-ais.com/article-images-pixabay/video-upscaling-consistency-optical-flow-7832555c.jpg)
framemeister xrgb mini back

## Options compared: flow vs LPIPS vs hybrid

Three approaches to 4× video super-resolution make different tradeoffs. A flow-based method, such as a BasicVSR++-style model using bidirectional propagation, is strongest when the priority is accurate motion alignment and low optical-flow endpoint error. An LPIPS-trained method is optimized for perceptual similarity and is the practical default when the output must look convincing to viewers. A hybrid combines flow-based alignment with LPIPS-guided refinement, but the available sources provides no direct head-to-head preference result for that combination.

Checkpoint 2 tests static-region flicker. Compute frame-to-frame luminance variance within the same static-region mask for every output sequence. The observed flicker energy is 100 units for the flow-based model and 115 units for the LPIPS-trained model. The latter is 15% above baseline because (115 − 100) ÷ 100 × 100 = 15%. That is below the rejection gate, which applies only when flicker energy exceeds the flow-based baseline by more than 20%; at this input, the equivalent cutoff is 120 units.

Checkpoint 3 is the perceptual decision. Measure LPIPS on the 4K outputs and enter both values in the worksheet. Begin with whichever candidate has the lower LPIPS, because perceptual quality is the default selector at the target scale. Do not reject that candidate unless its flicker energy is more than 20% above the flow-based baseline or its EPE exceeds 1.0 pixel. Here, EPE and flicker both remain within their gates, so the recorded LPIPS values determine the selection. This worksheet therefore converts the three measurements into one repeatable decision rather than treating temporal motion accuracy as a substitute for perceptual quality.

![Options compared: flow vs LPIPS vs hybrid — Video Upscaling Consistency](https://static.mm-ais.com/article-images-pixabay/video-upscaling-consistency-optical-flow-70f2e3b5.jpg)

## Decision rules: five if/then calls

**If** the LPIPS-trained model has the lowest LPIPS score, its EPE is below 1.0 pixel on your own footage, and its flicker energy is no more than the flow-based baseline plus 20%, **then** ship the LPIPS-trained model. This is the default call: perceptual quality leads the decision, provided that the model has not introduced unacceptable alignment error or temporal shimmer.

**If** the candidate’s EPE exceeds 1.0 pixel, **then** switch to flow-based propagation regardless of its LPIPS score. Alignment errors can undermine otherwise convincing frames, so a perceptual advantage does not override motion failure.

**If** the LPIPS-trained model’s flicker energy exceeds the flow-based baseline by more than 20%, **then** reject that model even if it wins a human-preference test. Temporal shimmer can mask perceived gains across consecutive frames, making the apparent quality advantage less useful in sustained playback.

**If** the lowest-LPIPS model fails either guardrail, **then** select a model that passes both EPE and flicker checks; do not average the scores or treat one metric as compensation for another. If no candidate passes both, do not ship on perception alone. Test a hybrid or revised configuration and measure the same three quantities again.

**If** EPE is exactly 1.0 pixel or flicker is exactly the permitted baseline-plus-20% limit, **then** treat the candidate as conditionally acceptable rather than automatically rejected. The stated limits are pass/fail boundaries, so confirm the result on a representative sequence and document any visible artifact before release.

**If** a new model changes the comparison, **then** recompute the flow-based baseline, recalculate the 20% flicker allowance, and rerun EPE and LPIPS measurements under the same target scale and evaluation footage. A stale baseline can turn an otherwise valid rejection into the wrong deployment decision.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | Evaluate every candidate at the same target scale, using 1080 as the common comparison point. | LPIPS is a valid default selector only when the models are compared at the same target scale. |
| 2 | Rank the evaluated models by LPIPS and start with the model having the lowest score. | Perceptual similarity is the default decision criterion; flow-based models cannot be assumed to win on quality. |
| 3 | Measure optical-flow endpoint error (EPE) on your own footage and reject the leading model if EPE is above 1.0 pixel. | This checks whether temporal alignment is strong enough on the content you will actually use. |
| 4 | Compare the leading model’s flicker energy with the flow-based baseline, applying the 20% rejection threshold. | Reject the lowest-LPIPS model only if its flicker energy exceeds the baseline by more than the defined threshold. |
| 5 | Review beta-FFT separately from the flow-based model rather than treating either approach as automatically superior. | beta-FFT uses a nonlinear method, while the available sources reports perceptual inferiority for flow-based models with an 18.4% preference figure. |
| 6 | Select the model that retains the lowest LPIPS without violating either the EPE or flicker-energy threshold, and document the three measurements together. | Keeping EPE, flicker energy, and LPIPS visible makes the 2026 selection rule auditable rather than preference-driven. |

## Frequently Asked Questions

**Which three metrics should be used to evaluate video-upscaling consistency?**

Use optical-flow endpoint error (EPE) for temporal alignment, flicker energy for temporal instability, and LPIPS for perceptual similarity.

**Which model should be selected first when comparing models at the same target scale?**

Select the model with the lowest LPIPS score at the target scale.

**When should a model with the best LPIPS score be rejected?**

Reject it if its flicker energy exceeds the flow-based baseline by more than 20% or if its EPE is above 1.0 pixel on the footage.

**What does optical-flow endpoint error measure in an upscaled video?**

EPE measures motion alignment by comparing the flow estimated from an upscaled frame with the flow estimated from the reference frame.

**Should a flow-based model be assumed to provide better perceptual quality?**

No; the available sources report that flow-based models are perceptually inferior, with an 18.4% preference rate.

**What consistency technique does beta-FFT use?**

Beta-FFT uses a nonlinear interpolation consistency loss across interpolated data points.

## Quick answers

| Which three metrics should be used to evaluate video-upscaling consistency? | Use optical-flow endpoint error (EPE), flicker energy, and LPIPS to assess temporal alignment, temporal instability, and perceptual similarity, respectively. |
| --- | --- |
| Which model should be selected first when comparing models at the same target scale? | Start with the model that has the lowest LPIPS score at the target scale, because LPIPS is the default selection criterion. |
| When should the model with the lowest LPIPS score be rejected? | Reject it if its flicker energy exceeds the flow-based baseline by more than the 20% threshold or if its EPE is above the 1.0-pixel threshold on the footage. |
| What does optical-flow endpoint error measure? | EPE measures motion alignment by comparing the flow estimated from an upscaled frame with the flow estimated from the reference frame. |
| Does a flow-based model necessarily provide the best perceptual quality? | No, the available sources report that flow-based models are perceptually inferior with an 18.4% preference rate, while beta-FFT uses a nonlinear interpolation consistency loss across interpolated data points. |

Also worth reading: **What to expect from 7900 XTX for 4K video upscaling**: [What to expect from 7900](https://ai-videoupscale.com/blog/what_to_expect_from_7900_xtx_for_4k_video_upscaling.php) · **Stop video flicker in 8K: 2026 RIFE v3.0 vs FILM at 42dB Peak Signal-to-Noise (PSNR-HVS-M)**: [Stop video flicker in 8K:](https://ai-videoupscale.com/blog/stop-video-flicker-in-8k-2026-rife-v30-vs-film-at-42db-peak-signal-to-noise-psnr-hvs-m.php) · **Upscale video to 4K: Basic Video Super-Resolution++ propagation vs. per-frame at 2×**: [Upscale video to 4K: Basic](https://ai-videoupscale.com/blog/upscale-video-to-4k-basic-video-super-resolution-propagation-vs-per-frame-at-2.php)

### Related reading

- [Temporal Consistency Loss: 4K Flicker Reduced 30% (LPIPS/VMAF)](https://ai-videoupscale.com/blog/temporal-consistency-loss-4k-flicker-reduced-30-lpipsvmaf.php)
- [NVIDIA RMA Process Analyzing the Quality and Consistency of Refurbished Graphics Card Replacements](https://ai-videoupscale.com/blog/nvidia_rma_process_analyzing_the_quality_and_consistency_of.php)
- [Quality Considerations When Converting AVI to MP4 A Technical Guide to Upscaling Results](https://ai-videoupscale.com/blog/quality_considerations_when_converting_avi_to_mp4_a_technica.php)
- [Exploring the Intersection of Optical Art and AI A 2024 Perspective on Digital Op Art Upscaling](https://ai-videoupscale.com/blog/exploring_the_intersection_of_optical_art_and_ai_a_2024_pers.php)
- [Step-by-Step Guide Extracting and Upscaling High-Quality Still Frames from Magix Movie Edit Pro 2025](https://ai-videoupscale.com/blog/step_by_step_guide_extracting_and_upscaling_high_quality_sti.php)
- [How to Convert MKV Files for Premiere Pro While Preserving AI Upscaling Quality](https://ai-videoupscale.com/blog/how_to_convert_mkv_files_for_premiere_pro_while_preserving_a.php)

### Latest

- [Facebook video upload quality 2026: 1080p vs 4K retention (VMAF)](https://ai-videoupscale.com/blog/facebook-video-upload-quality-2026-1080p-vs-4k-retention-vmaf.php)
- [Upscaling Puzzle Pirates 800x600 to 4K: Temporal Consistency Stops 2026 Flicker](https://ai-videoupscale.com/blog/upscaling-puzzle-pirates-800x600-to-4k-temporal-consistency-stops-2026-flicker.php)
- [How Weta Digital Upscaled X-Men: First Class to 4K: 2-Frame Temporal Window vs 5](https://ai-videoupscale.com/blog/how-weta-digital-upscaled-x-men-first-class-to-4k-2-frame-temporal-window-vs-5.php)

Canonical: https://ai-videoupscale.com/blog/video-upscaling-consistency-optical-flow-error-flicker-quality3-metrics-for-2026.php
Markdown: https://ai-videoupscale.com/blog/video-upscaling-consistency-optical-flow-error-flicker-quality3-metrics-for-2026.php/index.md