QBVE: QUADRATIC BAD VMAF ENERGY

A quality metric that exposes what VMAF averages hide

A video can reach an "excellent" mean VMAF of 94 and still contain seconds of visible quality collapse. QBVE, developed at Wavelet Beam, measures only the energy of these failures below a quality floor. It turns a subjective glitch into a hard, measurable value and makes encoder comparisons honest.

The Problem with Averages

Arithmetic and harmonic means blend quality drops with high-performing frames. A severe glitch in a high-motion scene is statistically masked by static, perfect scenes. Automated convex hull calculations based on such averages select bitrate points that look good on paper but fail in reality.

Beyond "1% Low" and Min VMAF: Quality Pumping

The industry has started to look at "1% low frame" statistics to spot quality drops. This approach is incomplete: a single low value such as the minimum VMAF only shows how deep a dip was. It does not capture the frequency, the duration and the temporal fatigue caused by constant quality fluctuations. We call this visual instability quality pumping, and traditional metrics cannot quantify it. QBVE measures the total energy of all failures below the quality gate, so both the intensity and the duration of every drop count. It quantifies the visual calm of a bitstream.

How QBVE Works

• Penalty filter: Frames above the quality floor contribute zero. Excellent frames cannot "save" the score of a failing frame.
• Quadratic penalty: The deviation below the floor is squared, so visible artifacts weigh far more than minor fluctuations. A dip to VMAF 75 is penalized about 5 times harder than a dip to VMAF 85.
• Gated MSE: QBVE works like a mean squared error that only opens below the gate. Lower is better, 0 means no frame fell below the floor.

Choosing the Threshold

The threshold T is the VMAF score below which artifacts become noticeable to the viewer and the QBVE penalty starts.

• UHD baseline: For high-end UHD content (3840x2160 at 50 fps) in HEVC we use a threshold of 93. Any drop below 93 is a measurable impairment of visual calm.
• 1080p: Lower resolutions show different spatial artifact perception. Using a linear interpolation based on the pixel count relative to UHD, we assume an initial threshold of 72 for 1080p (HEVC 1920x1080 at 50 fps). As a rule: the lower the video resolution, the lower the threshold.
• Comparability: A formal normalization across resolution tiers is possible but not required, as long as QBVE is evaluated consistently within the same test series.

Worked Example: When 1 VMAF Point Means Everything

A 10 minute video at 50 fps has 30,000 frames and a mean VMAF of 94. If about 1,364 frames, roughly 27 seconds, collapse to VMAF 72, the mean drops by just 1 point: (1,364 × 22) / 30,000 = 1. With a floor of 93, QBVE reports (1,364 × 21²) / 30,000 = 20.05. A 1 point drop sounds harmless, a QBVE of 20.05 shows that it is not.

Case Study: Standard Encoding vs. IRIS (UHD)

Netflix test content "Meridian", 43,092 frames in UHD (3840x2160), encoded in HEVC at comparable bitrates, measured with IRIS.ANALYST. Without IRIS the signal is highly volatile: even with a high mean, frequent quality dips result in a QBVE of 9.465, a restless, flickering image. With IRIS-Denoising and per-shot encoding, the entropy is stabilized before it reaches the encoder.

The minimum VMAF improved by less than 2 points, but QBVE was nearly halved at just 3% more bitrate. IRIS does not only raise the floor, it removes the constant pumping of quality and delivers a rock-solid UHD experience.

Quality analysis per frame: without IRIS (left, QBVE 9.465) and with IRIS (right, QBVE 5.062).

Choosing the Right Reference

Measuring VMAF, PSNR and QBVE when noise reduction is involved creates a benchmarking paradox: standard metrics penalize any deviation from the source, including the removal of unwanted sensor noise. We therefore measure:

• IRIS workflow: denoised source vs. encode with IRIS-Denoising
• Standard workflow: original source vs. encode without IRIS-Denoising

Against the noisy original, PSNR of a denoised encode drops into the 30 dB range, simply because the removed noise counts as an error. VMAF on the other hand is largely blind to high-frequency random noise: a denoised file against the original scores between 98 and 100. That proves IRIS-Denoising does not degrade the actual picture content, but it says nothing about how well the encoder preserves the cleaned signal. The denoised file is therefore the ground truth for all IRIS workflows. Conceptually, it reflects what an observer standing next to the camera on set would have seen: the scene without sensor noise and without compression artifacts.

A Second Problem: No Error Intervals

VMAF is a machine learning model trained on subjective viewer ratings, and every such model has prediction uncertainty. Yet no systematic error intervals have been published for VMAF scores. When one encoder scores 94.2 and another 93.8, nobody can tell whether the difference is significant or within the model's prediction noise. QBVE addresses the first problem directly and makes the second one impossible to ignore.

QBVE in Our Workflow

• IRIS and IRIS.ANALYST: IRIS cleans the image, IRIS.ANALYST generates the metadata that drives the encoding strategy.
• Per-shot encoding with FFmpeg: the analyst data optimizes the bitrate distribution using industry-standard tools, without a proprietary encoder.
• QBVE measurement: once the HLS files are generated, our QBVE tool calculates the final stability score. Bitrate savings must not be bought at the expense of visual consistency.

IRIS.ANALYST and the QBVE tool are available to Wavelet Beam customers. As a complementary check, we also count frames below VMAF 93, below VMAF 87 and below 40 dB PSNR.

Stop measuring averages, start eliminating failure. Want to see what QBVE reveals in your most challenging content? Contact info@waveletbeam.com for a QBVE analysis.