VMAF, PSNR and SSIM are quality metrics used in encoding and research, not universal verdicts. They compare a test video to a reference and produce a number. A high score can be useful evidence, but it does not guarantee that every viewer will prefer the file.
FAST-VID does not calculate VMAF, PSNR or SSIM and does not provide a quality-testing suite. This guide explains what the metrics mean for people evaluating encoder output elsewhere.
Reference-based comparison
These metrics need a reference video and a processed version. The tool aligns the frames, compares them and aggregates a score. If the reference is wrong, cropped, shifted, scaled or encoded differently, the result may not describe the quality you care about.
A reference comparison is not the same as judging a standalone stream. Without the original, a metric cannot tell you what detail was lost or whether the source was already soft.
The reference must represent the intended original. If it is already compressed, scaled or cropped, the score describes a different comparison.
Frame alignment, color range and timing can change the result. A small offset can make a good encode look worse.
A metric cannot recover a missing original. If the reference is not the intended master, the score answers the wrong question.
The reference and test video should share resolution, frame rate and timing; otherwise the comparison may measure alignment more than encoding quality.
What PSNR measures
PSNR compares pixel differences and expresses the result on a logarithmic scale. It is simple and widely used, but it treats errors in a way that does not always match human perception. A higher PSNR generally means a lower measured pixel error, not necessarily a more pleasing picture.
PSNR can be useful for detecting gross differences or comparing encodes under controlled conditions. It can also reward blur that reduces pixel error while making detail look worse.
PSNR is easy to compute and explain, but it does not model how people see edges, texture or motion.
A blurred file can reduce pixel error in some scenes while looking less detailed, which is a known limitation of simple error metrics.
PSNR can be useful for controlled regression tests, but it should not be the only evidence of visual quality.
A high PSNR on a soft encode can coexist with lost texture and detail.
What SSIM measures
SSIM looks at luminance, contrast and structure rather than raw pixel error alone. It was designed to correlate better with perceived quality than PSNR, but it is still a model with assumptions, scales and implementation details.
SSIM values and variants are not all interchangeable. A score from one implementation or spatial window may not be directly comparable to another.
SSIM uses local statistics and can be more perceptually relevant than PSNR, but its score depends on the window, scale and implementation.
Different SSIM variants are not all equivalent; compare like with like.
SSIM can be more sensitive to structure, but it still uses a limited model of human vision.
Averaging local scores can hide a severe artifact in a small but important region.
What VMAF measures
VMAF is a perceptual metric developed by Netflix. It uses features and a trained model to estimate quality relative to a reference. It is widely used in streaming and encoding research, but it remains a model, not a direct measurement of every viewer's experience.
VMAF versions, models, scaling and viewing conditions matter. A high VMAF score is useful evidence, but it does not prove that a file is universally better for every display, content type or viewer.
VMAF was designed for video quality assessment and is widely used in encoding research and streaming.
Model version, training data and viewing conditions affect what a VMAF score means.
VMAF is useful for comparing encodes on similar content and settings, not for declaring one codec universally superior.
Test conditions and metric limitations
Content type, resolution, frame rate, bit depth, color and viewing distance affect how errors are perceived. A metric tuned for one condition may be less reliable in another. Graphics, animation, film grain and dark scenes can stress different parts of the model.
Metrics can also be gamed by encoders that optimize for the score. A higher number can hide artifacts in a region the metric weights less heavily. Human viewing remains necessary for final decisions.
A metric can be optimized by an encoder without producing a universally better picture.
Dark scenes, text, film grain, animation and fast motion can stress metrics differently.
Scores at different resolutions or frame rates are not automatically comparable.
A metric may favor one kind of artifact over another; human review can reveal what the score misses.
Viewing conditions and display calibration change perceived quality, so a score from one setup is not universal.
Metrics are useful for ranking candidates under controlled conditions, but a single number should not override visible artifacts.
| Metric | Main focus | Key limitation |
|---|---|---|
| PSNR | Pixel error | Weak correlation with perception |
| SSIM | Local structure and contrast | Model and implementation dependent |
| VMAF | Perceptual model against a reference | Model version and test conditions matter |
How to interpret a score
Use a metric as one input, not a verdict. Compare files under the same conditions, keep the reference and settings consistent, and inspect representative scenes at the intended viewing size. Do not compare scores across codecs, resolutions or metric versions as if they were the same scale.
FAST-VID does not calculate these scores or offer quality testing. It downloads available source formats. Only use content you own or are authorized to save.
Use metrics alongside human review at the intended screen size and viewing distance.
Report the metric version, reference and settings when comparing results.
FAST-VID does not calculate quality scores; the numbers belong to your own testing workflow.
Use a metric as a screening tool, then inspect representative frames and motion at the intended size.
When scores disagree with your eyes, investigate the specific scene, metric version and test setup rather than assuming the numbers are wrong.
Document the reference, metric version, resolution, frame rate and viewing setup; a score without context is hard to reproduce.
Frequently asked questions
What is VMAF?
VMAF is a perceptual video quality metric that compares a test video with a reference. It is a model, not a universal guarantee of viewer preference.
Is a higher PSNR always better?
No. PSNR measures pixel error, which does not always match perceived quality. Blur or other changes can produce a higher score while losing detail.
What does SSIM measure?
SSIM compares luminance, contrast and structure. It was designed to correlate with perceived quality, but it is still a model with implementation and content limits.
Can a score prove a file is better for everyone?
No. Scores depend on the reference, metric version, content, scaling and viewing conditions. Human viewing is still needed for a final judgment.
Does FAST-VID calculate VMAF, PSNR or SSIM?
No. FAST-VID is a public-link downloader and does not provide quality metrics, encoding tests or video-quality scoring.
Sources and references
- Netflix VMAF repository
- ITU-T P.910: Subjective video quality assessment
- Wang et al.: Image quality assessment using SSIM
- FAST-VID: Video compression explained
Try FAST-VID
Open the FAST-VID downloader to inspect available source formats. FAST-VID does not calculate VMAF, PSNR or SSIM and does not provide a quality-scoring feature.