How we score video quality — and what VMAF actually measures

The short answer

A single "quality +30" score often mixes two jobs: repair, which cleans the picture at the same size, and upscaling, which adds pixels; stretching 480p to 4K can raise it without cleaning anything. After a preview, KwaFlux quality detection reports three parts instead: source quality, same-resolution repair, and resolution delivery. VMAF measures fidelity to a reference, not a glow-up.

The single-score trick

Tool A blows a soft 480p clip to 4K and barely denoises. Tool B outputs 1080p but actually removes blocks. A before/after total often crowns A, because resolution is the cheapest way to look improved. Six months later you are watching B's work. Resolution is easy to score. Repair is the work. One number systematically rewards the first and hides the second.

Three numbers, three questions

1. Source quality — sharpness, cleanliness, color, motion on the original. Where did you start, and how much headroom is left?
2. Same-resolution repair — compare processed vs original at one size. Denoise, deblock, and texture fidelity without the canvas jump
3. Resolution delivery — the upscale gain, gated by repair quality. A dirty frame grown larger is a larger dirty frame

The delivery total is repair content plus that gated resolution term. All three stay visible. We describe the idea, not internal field names.

VMAF is fidelity, not a glow-up

• VMAF (Netflix, open source) answers "how like the reference is this frame?" — not "how much better did we make it?"
• High VMAF = close to the reference style. Low VMAF = a large rewrite, which might be an improvement or over-cooking. Eyes decide
• If a vendor shouts "VMAF up X%", ask what the reference was. If it is a naive stretch of the bad original, a high score can mean "we changed little"

Inside KwaFlux, VMAF is a fidelity hint that flags a heavy rewrite for a visual check. It is not advertised as a quality-up proof. Detection is a navigation aid across settings. Download and settle the 1s/3s/5s preview yourself.

Use the split as a map, not a verdict

The three-part report exists so a resolution jump cannot hide a weak repair. It is still not a grade you should quote as proof that the picture improved. Source quality says where the clip started. Same-resolution repair says whether denoise, deblock, and texture held when the canvas did not change. Resolution delivery is the upscale term, and it is gated by that repair. A dirty frame grown larger is a larger dirty frame. If the delivery total looks flattering while the repair line barely moved, you are watching the single-score trick from the inside. Read the split.

VMAF does not change that reading. It answers a narrower question: how close is this frame to a reference you chose. Netflix published it as an open fidelity metric. A high value means close to that reference. A low value means a large rewrite. Close can mean the pass barely touched a damaged original. Far can be a cleaner picture or an over-cooked one. Your eyes on the preview decide which. If a vendor prints a VMAF rise as a glow-up, ask what the reference was. A plain stretch of the damaged source can score as faithful because little changed. That figure is not a repair, and we do not publish a VMAF number for KwaFlux output.

• Read repair before delivery. If repair is flat, do not treat the total as the story
• Keep the reference explicit. Fidelity to the original is a different claim from fidelity to a clean master you do not have
• Treat a sharp VMAF drop as a flag to look, not as a pass or a fail. A heavy denoise rewrites texture. That can be the point, or it can be plastic
• Compare two settings on the same 5s beat. The report navigates those settings. It does not sign the master
• Do not relabel our detection output as a VMAF score. Detection suggests where to look first

Download KwaFlux and run the 1s/3s/5s preview on your own damaged footage. Those previews need no account; export is on pricing. The useful habit is simple: let the breakdown tell you whether the processing time went into cleaning or into a bigger canvas, then let your eyes approve the frame. A single score that mixes both jobs is still a mix. We keep the jobs separate so the easy one can't pass for the hard one. One more trap: averaging the three lines into one story for a client. The lines can disagree. Strong repair with a modest delivery score is a cleaned picture that stayed the same size. Weak repair with a big delivery score is a larger copy of the same damage. Report the disagreement. Don't collapse it into a single gain you didn't measure, and don't borrow a VMAF figure to dress it up.