Comparing CRF Settings — Measure Encoder Quality, Don't Guess

The CRF folklore problem

Every encoding guide repeats the same folklore: “CRF 18 is visually lossless, CRF 23 is fine for the web.” Sometimes true — but CRF is content-dependent. A talking-head clip at CRF 26 can look pristine while a confetti-filled concert shot falls apart at CRF 20. The only honest answer for your footage is a measurement.

The two-encode test

1. Take a short, representative slice of your footage — include the hardest content you have (motion, grain, fine detail). A 20–30 second cut is ideal.
2. Encode it at the two candidate settings, e.g. CRF 20 and CRF 24.
3. Compare each encode against the original with DiffALL: upload original + encode, read the SSIM score and the per-second chart.

If CRF 24 scores 97%+ SSIM against the source on your hardest section, the extra bitrate of CRF 20 is buying you nothing you can see — ship the smaller file. If it dips to 90% during the action scene, you’ve found your reason to spend bits, and the chart shows exactly which seconds suffer.

Why per-second beats a single score

Encoders fail locally. An average SSIM of 96% can hide a two-second block-artifact disaster during the one fast pan in the clip. DiffALL scans every second, flags the worst moments, and charts quality over time — so you judge an encode by its weakest seconds, which is exactly how viewers do.

Reading the numbers

• SSIM 97%+ — differences are essentially invisible at normal viewing distance.
• SSIM 93–97% — fine for most web delivery; inspect the flagged seconds.
• Below 90% — visible degradation; raise the bitrate or lower CRF for this content.

The same method settles any encoder question, not just CRF: x264 vs x265 at equal bitrate, preset slow vs fast, hardware vs software encoding, or whether your streaming platform’s re-encode mangled your upload — compare what you uploaded against what it plays back.

Keep clips short

Comparison on the free plan covers clips up to 1 minute and 40 MB (90 seconds and 150 MB with Pro) — which fits this workflow naturally: you want a short, hard test slice anyway, not a full episode.

Try it

Cut a test slice, encode two candidates, and measure them — SSIM, PSNR, per-second chart, and a synchronized side-by-side player. Free, in the browser.