MSU VQMT, FFMetrics and VMAF GUI Tools Compared (2026)

Updated November 2026: VMAF v1 (libvmaf 3.2.0) versus MSU VQMT 14.1, the rebuilt Elecard Video Quality Estimator, and which tools give VMAF a GUI.

Why this matters

FFmpeg with the libvmaf filter gives you a number; a dedicated tool gives you an investigation. That difference matters the moment a clip fails a quality check and someone asks which frame, and caused by what. This article is for the engineer who already computes VMAF in a pipeline — a streaming or encoding lead, a QA engineer, a technical product owner — and now needs to look at quality, compare many encodes interactively, score HDR content, or hand a non-command-line colleague a tool they can drive. Pick the right dedicated tool and you find the failing frame in minutes. Pick the wrong one — or read its output on the wrong scale — and you ship a confident conclusion built on a misread plot.

A dedicated tool is a workbench, not a new metric

Start with the distinction that saves the most confusion. A metric is the math that estimates quality — the perceptual score called VMAF, the pixel-error measure called PSNR (Peak Signal-to-Noise Ratio, in decibels), the structural measure called SSIM (Structural Similarity, on a 0–1 scale). A tool is the program that computes that math and hands it to you. The deep dives on each metric live in VMAF explained and the rest of Block 2; this article is about the tools.

The point that trips people up: a dedicated tool does not give you a better metric. The VMAF that MSU VQMT computes is the same VMAF the FFmpeg-and-libvmaf workflow computes, because both call the same model. What a dedicated tool adds is everything around the number: more metrics computed in one pass, a plot of the score across every frame, the ability to click a low point and see the actual frame, a heatmap of where the two videos differ, GPU acceleration, and a graphical interface a non-engineer can use. Think of FFmpeg as a precise digital scale that prints a weight, and a dedicated tool as the full lab bench — the same scale, plus the magnifier, the calipers, and the chart recorder beside it.

First, untangle the name: there are two "VQMT"s

Before going further, clear up a collision that wastes afternoons. The string "VQMT" — Video Quality Measurement Tool — names two different programs from two different authors.

The one this article means is MSU VQMT, a large Windows and Linux application from the Moscow State University Graphics & Media Lab Video Group and the COMPRESSION.RU team (project lead: Dr. Dmitry Vatolin). It computes roughly two dozen metrics, has a full graphical interface, GPU acceleration, HDR metrics, and paid Pro and Premium tiers (MSU VQMT documentation, version 14, 2026).

The other is rolinh/VQMT, a small open-source command-line program written in C++ on top of OpenCV. It computes six classic full-reference metrics — PSNR, SSIM, MS-SSIM, VIFp, PSNR-HVS, and PSNR-HVS-M — from raw video, and it does not compute VMAF (rolinh/VQMT, GitHub). It is free and scriptable, but it is not the feature-rich application most articles mean when they write "VQMT." When you read a tutorial, check which one it is by whether VMAF and a GUI are on the table. If they are, it is the MSU tool.

Check the version before you plan around it. The latest public MSU VQMT release is still 14.1, from June 2022, and the developers’ changelog lists nothing newer as of November 2026 (MSU VQMT changelog, checked November 2026). That makes it older than VMAF v1, which Netflix first shipped in libvmaf 3.2.0 on June 20, 2026 (Netflix/vmaf releases, 2026). If your gate has moved to VMAF v1, compute that score with libvmaf and use VQMT for the investigation around it. Do not set a VQMT VMAF beside a v1 number and read the gap as a quality change.

What MSU VQMT adds over FFmpeg

Take the dedicated tool one capability at a time, because each answers a different "why would I leave FFmpeg" question.

More metrics in one run. MSU VQMT computes PSNR, MSE, MSAD, Delta and Delta ICtCp, SSIM, MS-SSIM, 3SSIM, the NTIA Video Quality Metric (VQM), VMAF, the no-reference NIQE, and the spatial- and temporal-information measures SI and TI, plus MSU's own no-reference blocking and blurring detectors — alongside HDR variants of several of these (MSU VQMT documentation, 2026). FFmpeg gives you PSNR, SSIM, XPSNR and VMAF, with MS-SSIM available through libvmaf; VQMT gives you those plus a wider bench when you want a second or third opinion on the same pair of files.

It scores HDR correctly. Standard-dynamic-range metrics misread high-dynamic-range video, because HDR uses a different transfer function and color space with very different brightness levels, so an SDR metric does not "see" the pixels the way an HDR display would. VQMT added HDR-specific PSNR, SSIM, MS-SSIM, VQM, and Delta ICtCp for exactly this reason, and lets you set the input color space and simulate a display color space (MSU VQMT version 13 notes, via Streaming Learning Center, 2022). VMAF still has no official HDR model: the VMAF v1 release in libvmaf 3.2.0 adds SDR models only (Netflix/vmaf releases, June 2026), so this is a real gap a dedicated tool fills. HDR needs a paid license, though. Since version 14.1 the Free edition leaves out HDR files, HDR metrics and GPU metrics (MSU VQMT changelog, 2022).

It runs on the GPU. VQMT implements its SSIM-family metrics on the GPU through OpenCL and CUDA, and accelerated VMAF through OpenCL that works on both AMD and NVIDIA cards. In one independent test, GPU acceleration computed VMAF 40–67% faster than the CPU path on the same machine, with identical scores (Jan Ozer, Streaming Learning Center, 2022).

It works in batches, regions, and three interfaces. VQMT runs as a Windows GUI (free and paid), a Windows-and-Linux command line (paid Pro tier), and a Python package, so the same engine serves both an interactive session and an automated job. It computes several metrics simultaneously, restricts measurement to a region of interest, exports to CSV or JSON, and can automatically save the "bad" frames where two encodes differ most — the frames you actually want to look at (MSU VQMT documentation, 2026).

Figure 1. What a dedicated tool adds. FFmpeg computes the metric and prints a number; MSU VQMT computes the same metric and adds the workbench around it — more metrics, per-frame plots, frame inspection, a difference heatmap, GPU acceleration, and HDR-aware variants.

The capability that matters most: seeing the frame that failed

If a dedicated tool earns its place for one reason, this is it. FFmpeg's libvmaf can write a per-frame log, but you still have to plot it yourself and then go hunting for the matching frame. VQMT closes that loop inside one window. It draws a results plot — the quality score of one or two encodes across the whole clip — so a dip is visible at a glance. You move the playhead to the dip and show the source frame and the encoded frame side by side. Then a difference (residue) view paints where the two frames disagree, with the brighter areas marking larger distortion, so the artifact is not just a number on a chart but a region you can point at (MSU VQMT, Streaming Learning Center, 2022).

This is the daily reason engineers keep a dedicated tool next to their FFmpeg scripts. A single pooled VMAF score — the average across all frames — is a summary, and the mean is the summary most likely to hide a short, severe problem. Remember the rule from pooling per-frame scores: the average flatters, and a one-second collapse barely moves it. The per-frame plot is how you catch the collapse the average erased.

Figure 2. The find-the-bad-frame loop. A per-frame plot exposes a dip the mean hides; you jump to that frame, compare source and encode, and a difference heatmap localizes the artifact — the investigation a bare pooled number cannot give you.

A worked example makes the gap concrete. Suppose a 10-minute clip — 600 one-second samples — returns a mean VMAF of about 96, comfortably "good." The per-frame plot, though, shows the score holding near 96 for the whole clip except a two-second window where it falls into the low 70s (73, then 71) during a hard cut. The mean barely noticed, and here is the sharper lesson: the dip is so short that even the 5th-percentile score misses it.

Mean VMAF      = (96 × 598 s + 73 + 71) ÷ 600 s = 95.9   ← still "good"
5th-percentile = 96    ← a 2-second dip is too short (≪ 5% of 600 s) to move it
Minimum sample = 71    ← only the minimum, and the plot, expose the problem

This is the case for a per-frame visual tool, not just better pooling. Reading the low percentile and the minimum instead of the mean is the right habit — the discipline in reading a quality-metric report — but a short, severe dip can slip below even the percentile. The plot does not average anything away: the collapse is a visible notch you click straight to. That one-click jump to the dip is the capability you are buying.

When the speed knob is worth turning

Dedicated tools also expose performance controls FFmpeg hides. VQMT's subsampled mode scores every Nth frame instead of all of them. In one test, scoring two one-minute files on every frame took 131 seconds; scoring every fourth frame dropped that to 42 seconds — a 68% reduction — while neither the VMAF mean nor the low-frame score changed by a meaningful amount (Streaming Learning Center, 2022):

Full scan:        131 s  (every frame)
Every 4th frame:   42 s  → (131 − 42) ÷ 131 = 68% less time (≈3.1× faster)
Mean VMAF, low-frame VMAF: essentially unchanged

That is a useful trade when you are triaging a large catalog and want a fast first pass, with the caveat to confirm on your own content before trusting it, because subsampling can skip the exact frame you needed to see. Speed via subsampling, threads, or the GPU is a convenience; it never changes which metric or model produced the score — and the version still matters. Netflix calibrated VMAF v1 to line up with v0 scores, but v1 adds new features and turns NEG mode on by default (Netflix Technology Blog, June 2026; see VMAF NEG explained), so moving from v0 to v1 can still shift individual clip scores. Record the version every dedicated tool reports, exactly as VMAF in depth insists.

The peers worth knowing

MSU VQMT is the broadest dedicated tool, but four neighbors solve specific problems better or cheaper.

Apple AVQT (Advanced Video Quality Tool) is a macOS command-line tool that returns a single perceptual score from 1 to 5 — the same scale as a Mean Opinion Score from human raters — for content with compression and scaling artifacts. It is full-reference (it needs the original), handles SDR and HDR through the AVFoundation framework, uses the Metal GPU to run faster than real time, and lets you set viewing distance, display resolution, segment duration, and the temporal pooling method, writing results as JSON or CSV (Apple, WWDC21/WWDC22). As of November 2026 Apple still offers it for macOS only, and those two sessions remain its developer documentation for the tool. Its 1-to-5 output is its trap: never compare an AVQT 4.2 to a VMAF 92 as if they sat on one scale. AVQT's place in the metric landscape is covered in beyond VMAF.

FFMetrics is the free shortcut to per-frame visualization. It is a Windows graphical front-end that runs FFmpeg under the hood and charts the PSNR, SSIM, VMAF, and XPSNR it computes, exporting averages and per-frame statistics to CSV and picking a VMAF model automatically (FFMetrics, GitHub). Version 1.7.0 is the current stable release, and the 1.8.0 beta adds a pixel-format control that converts the reference before each metric runs (FFMetrics releases, GitHub, checked November 2026). Look at which VMAF model it picked in the log before you compare its numbers with a VMAF v1 pipeline. If your only reason to leave the command line is "I want to see the plot," FFMetrics gives you that without a license — you supply your own ffmpeg.exe.

Elecard Video Quality Estimator is a commercial graphical tool that measures PSNR, SSIM, VMAF (including the phone model), VQM, VIF, and more, with per-frame graphs you can navigate, zoom into a chosen score range, and hover for a frame's exact value (Elecard, 2026). Elecard rebuilt the tool for StreamEye Studio 2026, released June 18, 2026: it now compares up to five files at once instead of three, opens HLS and MPEG-DASH manifests to score one profile across all segments, and installs on macOS without extra security settings (Elecard, 2026). It overlaps VQMT's interactive niche and is one option discussed in commercial quality suites.

rolinh/VQMT — the open-source namesake — is the lightweight pick when you want a few classic full-reference metrics computed fast from raw video with no GUI and no VMAF. Together with the broader blind-metric toolbox in open-source no-reference tools, it rounds out the free end of the bench.

Figure 3. The five dedicated tools at a glance: what each adds over FFmpeg, whether it needs a reference, its interface and cost, and where it lies.

The same comparison in text, for when you are choosing. The "where it lies" column is the limitation to keep in mind — every tool has one.

Table 1. Dedicated quality tools beside FFmpeg. The first column is the reason to add each; the "where it lies" column is the limitation. Names, tiers and versions re-verified against each tool's own documentation and release notes in November 2026.

Is there a GUI for VMAF?

Not from Netflix. VMAF ships as the libvmaf library, a command-line tool and the FFmpeg filter, so any VMAF GUI is a third-party front-end (Netflix/vmaf, GitHub). Three are worth your time. FFMetrics is free, runs on Windows and plots FFmpeg’s per-frame VMAF, PSNR, SSIM and XPSNR. MSU VQMT has a free Windows GUI, with HDR and GPU metrics in the paid tiers. Elecard Video Quality Estimator is commercial and, since the 2026 rebuild, installs cleanly on macOS as well.

The best tool for analyzing stream quality depends on what you need to see. For one automated number per encode, FFmpeg with libvmaf is enough. To look at the failing frame, pick FFMetrics if the budget is zero and you work on Windows, MSU VQMT if you need HDR metrics or a wide bench of metrics, and Apple AVQT if you work on a Mac and want one perceptual score on a 1-to-5 scale. For the wider field of tools, start with the video quality tooling landscape.

When the extra capability is worth it

The decision is not "VQMT versus FFmpeg" — most teams keep both. It is when to reach for the dedicated tool. Reach for it when you need to look at quality, not just record it: to find and show the frame that failed, to score HDR content where VMAF has no model, to compute metrics FFmpeg does not, to hand a graphical tool to someone who does not live in a terminal, or to push throughput up with the GPU. Stay on the FFmpeg-and-libvmaf path when you need a scriptable number inside an automated gate, where a GUI is dead weight and free-and-headless is exactly right. The fastest sort is: automated pipeline number, use FFmpeg; human investigation or HDR, use a dedicated tool.

Figure 4. When the extra capability is worth it. The choice branches on whether you need to see quality or only record it, whether the content is HDR, which operating system you run, and whether a paid tier is justified — not on how many files you have.

Common mistake: reading two tools' numbers on one scale. A VMAF from FFmpeg, a VMAF from MSU VQMT, and a 1-to-5 score from Apple AVQT can all look like "the quality number," but only the two VMAFs are comparable — and only if the model, the version, and the scaling match. AVQT's 1–5 is a different scale entirely; mapping a 4.2 onto VMAF's 0–100 is meaningless. Two tools can also disagree on the same metric if one upscaled the encode to the source resolution and the other did not, or if one ran VMAF v0 and the other v1. Before comparing, confirm both tools used the same metric, the same model and version, and the same scaling — and never compare a score across two metrics as if they shared an axis.

How to read their output

Dedicated tools speak in three forms, and each has a reading rule. The per-frame table (CSV or JSON) is the raw record — sort it to find the worst frames, and report the 5th-percentile and minimum next to the mean, never the mean alone. The plot is for spotting where quality moved over time; trust the shape, but read the y-axis units, because a curve drawn on a 90–95 axis exaggerates a half-point wobble into a cliff. The difference heatmap localizes distortion in space, with brighter regions marking larger error — a guide to where the artifact sits, not a calibrated score, so treat its colors as a pointer, not a measurement. Turning this output into clean, honest pictures for a report is its own craft, covered in visualizing quality.

Where Fora Soft fits in

Fora Soft has built video software since 2005 — streaming and OTT, video conferencing, e-learning, telemedicine, and surveillance — and our measurement bench has both kinds of tool on it. For the automated quality gate in a delivery pipeline we run FFmpeg with libvmaf, because it is free, headless, and scriptable. When an encode fails that gate, or a client asks why one preset looks softer than another, we open a dedicated tool to read the per-frame plot, jump to the failing frame, and show the difference heatmap so the answer is a picture, not an assertion. For HDR work we lean on the HDR-aware metrics a dedicated tool provides, since VMAF has no HDR model. The tool we recommend is the cheapest one that answers the actual question: FFmpeg for the number, a dedicated tool for the look. Our benchmark methodology documents exactly which tool, model, and version produced each figure. Quality gates like this are built into the pipelines our video processing software development team delivers, and our video troubleshooting and optimization service uses the same bench to explain why an encode looks wrong.

The handbook

Fora Soft Video Engineering Handbook

We wrote a seven-volume handbook that works through the whole video stack — encoding, delivery, audio and AI — so an engineer who reads it can design and build a pipeline and prove its quality with the right tool. Volume 1 walks through VMAF, PSNR and SSIM, the per-frame logs behind them and the encoder settings that move each score. Volume 3 takes the same measurement into ABR ladders, players and HDR delivery, and Volume 6 covers the learned models that newer perceptual metrics are built on. Written by Nikolay Sapunov, CEO at Fora Soft.

Browse the handbook on Amazon →7 volumes · Kindle & paperback

See our case studies

References

1. MSU Video Quality Measurement Tool (VQMT) — Basic information, metrics, editions, and features. MSU Graphics & Media Lab Video Group / COMPRESSION.RU Team (project lead Dr. Dmitry Vatolin); online help version 14, 2026. Tier 3 (first-party tooling documentation). The authoritative source for VQMT's metric list (PSNR, MSE, MSAD, Delta, Delta ICtCp, SSIM, MS-SSIM, 3SSIM, VQM, NIQE, VMAF, SI, TI, MSU blocking/blurring, HDR variants), its GPU acceleration (OpenCL/CUDA), region of interest, CSV/JSON export, bad-frame saving, the three interfaces (GUI, CLI, Python), and the Free/Pro/Premium/SDK editions. Basis for the "what MSU VQMT adds" section. https://videoprocessing.ai/vqmt/basic
2. VMAF — Video Multi-Method Assessment Fusion, Netflix, GitHub repository and releases (libvmaf v3.2.0, June 20, 2026: the first public release of VMAF v1, with v1.0.16 SDR models). Tier 1 (metric-author defining implementation). Documents that VMAF is computed by one shared model across the vmaf CLI, the libvmaf FFmpeg filter, and third-party tools — the basis for "a dedicated tool does not give you a better metric." https://github.com/Netflix/vmaf
3. Z. Wang, A. C. Bovik, H. R. Sheikh, E. P. Simoncelli, "Image Quality Assessment: From Error Visibility to Structural Similarity," IEEE Transactions on Image Processing, 13(4), 2004. Tier 1 (metric-author defining paper). The defining work for SSIM, one of the metrics every dedicated tool computes. https://ece.uwaterloo.ca/~z70wang/publications/ssim.html
4. Recommendation ITU-T P.910, "Subjective video quality assessment methods for multimedia applications," International Telecommunication Union, 2023. Tier 1 (official standard). Defines the spatial-information (SI) and temporal-information (TI) measures VQMT reports as no-reference content descriptors, and the subjective methods all objective metrics are validated against. https://www.itu.int/rec/T-REC-P.910
5. Evaluate videos with the Advanced Video Quality Tool (AVQT); What's new in AVQT. Apple, WWDC21 (session 10145) and WWDC22 (session 10149), Apple Developer. Tier 3 (first-party tooling). AVQT is a macOS command-line full-reference tool returning a 1–5 perceptual score for SDR and HDR via AVFoundation, with viewing-distance, display-resolution, segment-duration, and temporal-pooling flags, and JSON/CSV output. Basis for the AVQT entry. https://developer.apple.com/videos/play/wwdc2021/10145/
6. FFMetrics — Visualizes video quality metrics (PSNR, SSIM, XPSNR & VMAF) calculated by ffmpeg.exe. fifonik/FFMetrics, GitHub, 2026. Tier 6 (expert-practitioner tool). A free Windows GUI front-end that charts FFmpeg's per-frame metrics, exports CSV, and auto-selects the VMAF model. Basis for the FFMetrics entry. https://github.com/fifonik/FFMetrics
7. VQMT: Video Quality Measurement Tool. rolinh/VQMT, GitHub. Tier 6 (open-source tool). Fast C++/OpenCV command-line implementations of PSNR, SSIM, MS-SSIM, VIFp, PSNR-HVS, and PSNR-HVS-M from raw video; no VMAF, no GUI. Basis for the name-collision disambiguation. https://github.com/Rolinh/VQMT
8. J. Ozer, "MSU Updates Video Quality Measurement Tool," Streaming Learning Center, 2022. Tier 6 (expert practitioner). Independent account of VQMT's GPU/OpenCL VMAF acceleration (40–67% faster, identical scores), the version-12 subsampled mode (131 s → 42 s on every-fourth-frame), the version-13 HDR metrics and AV1 input, and the residue-plot visualization. Basis for the speed and HDR worked figures. https://streaminglearningcenter.com/metrics/msu-updates-video-quality-measurement-tool.html
9. Elecard Video Quality Estimator — product page and user guide, Elecard, 2026. Tier 6 (vendor). A commercial GUI tool measuring PSNR, APSNR, SSIM, Delta, MSE, MSAD, VQM, NQI, VMAF, VMAF phone, and VIF, with navigable, zoomable per-frame graphs and per-frame hover values. Basis for the Elecard entry. https://www.elecard.com/products/video-analysis/video-quality-estimator
10. "Calculating Video Quality Using NVIDIA GPUs and VMAF-CUDA," NVIDIA Technical Blog (with Netflix), 2024. Tier 4 (credible deployer). Reports the CUDA path for VMAF in FFmpeg and its throughput gains; corroborates the GPU-acceleration point for dedicated and command-line tools alike. https://developer.nvidia.com/blog/calculating-video-quality-using-nvidia-gpus-and-vmaf-cuda/
11. C. G. Bampis, Z. Li, et al., "VMAF v1: Good Is Not Good Enough," Netflix Technology Blog, 2026-06-20. Tier 1 (metric-author work). The June-2026 VMAF v1 release that makes recording the metric version mandatory when any tool reports a VMAF score. Basis for the version-discipline note. Published on the Netflix Technology Blog (Medium); the matching code release is libvmaf v3.2.0: https://github.com/Netflix/vmaf/releases/tag/v3.2.0
12. MSU VQMT — Version history & changelog. MSU Graphics & Media Lab Video Group, checked November 2026. Tier 3 (first-party tooling documentation). Lists 14.1 (June 2022) as the latest release and the Free-edition limits (no HDR files, HDR metrics or GPU metrics). Basis for the version and Free-tier notes. https://videoprocessing.ai/vqmt/changelog/
13. "StreamEye Studio 2026: Major Performance Boost with VMAF Acceleration and AV1 Buffer Support," Elecard, June 18, 2026. Tier 6 (vendor). The rebuilt Video Quality Estimator: up to five files at once, HLS and MPEG-DASH manifest analysis, Qt 6, macOS installation fix. Basis for the 2026 Elecard update. https://elecard.com/news/ses_2026
14. FFMetrics releases, fifonik/FFMetrics, GitHub, checked November 2026. Tier 6 (expert-practitioner tool). 1.7.0 stable and the 1.8.0 beta pixel-format control. https://github.com/fifonik/FFMetrics/releases