Video similarity metrics are numbers that show how similar a copy of a video is to the source: some compare the frames themselves, others compare digital frame fingerprints, and others keep an eye on picture quality. ClipClone calculates five such metrics for each copy: SSIM, pHash, dHash, PDQ and PSNR. Below are short definitions, what each number means and which thresholds the service uses.
What does each metric mean?
Each metric answers its own question: SSIM — is the structure of the frames similar; pHash, dHash and PDQ — do the digital frame fingerprints match; PSNR — is the picture degraded. A copy passes the check only if it stays within all thresholds at once.
SSIM
SSIM (structural similarity index) is a metric that compares two frames by brightness, contrast and structure and returns a number from 0 to 1. A value of 1.00 means the frames are identical; the lower the number, the more the copy differs from the source. In ClipClone, the SSIM of a copy with the source must be no higher than 0.88 in all built-in presets.
pHash
pHash (perceptual hash) is a digital frame fingerprint 64 bits long, built from the large details of the image rather than from individual pixels. Frames that look similar produce similar fingerprints. The difference is counted as the number of mismatched bits out of 64: 0 means the fingerprints match, more means the frames differ more. ClipClone threshold: at least 8 of 64 in all built-in presets.
dHash
dHash (difference hash) is a 64-bit frame fingerprint built from the brightness differences between neighboring areas of a downscaled image. It complements pHash: the two fingerprints react differently to changes in the frame. It is read in the same way — how many of the 64 bits did not match. ClipClone threshold: at least 8 of 64 in all built-in presets.
PDQ
PDQ is a perceptual image hash whose algorithm was open-sourced by Meta; in ClipClone its fingerprint is 256 bits long. The difference is how many of the 256 bits did not match. The algorithm's documentation recommends, as a starting point, treating fingerprints that differ by no more than 31 bits as matching (PDQ, ThreatExchange repository). The ClipClone threshold is stricter: at least 55 of 256 in all built-in presets.
PSNR
PSNR (peak signal-to-noise ratio) shows in decibels how much the picture of a copy is distorted compared with the source. In the ClipClone check it is a quality control, not a measure of difference: a PSNR that is too low means the changes have degraded the picture. The threshold in all presets is no lower than 12 dB. It catches only a truly degraded picture: any cropping, shift or color filter lowers PSNR to 14–17 dB without a visible loss of quality.
Hamming distance
Hamming distance is the number of positions in which two fingerprints of the same length differ. This is exactly how the difference by pHash, dHash and PDQ is counted: “a difference of 8 of 64” means that 8 of the 64 bits of the copy and the source did not match.
Median across frames
The median across frames is the value in the middle of the ordered series of per-frame differences; individual outliers barely affect it. ClipClone calculates fingerprints on 8 frames of the video and takes the median, so one random scene does not change the result of the check.
How do you read the numbers in the report?
Different metrics point in different directions: for SSIM, lower means a stronger difference; for pHash, dHash and PDQ, on the contrary, higher means a stronger difference; and PSNR must stay no lower than its threshold. An example of a real copy with the “Optimal” preset (V-04 on the home page):
| Metric | Value of the copy | “Optimal” threshold | What it means |
|---|---|---|---|
| SSIM with the source | 0.66 | no higher than 0.88 | the frame structure differs noticeably |
| pHash | 10 of 64 | at least 8 | the fingerprint differs more than the threshold requires |
| dHash | 11 of 64 | at least 8 | the fingerprint differs more than the threshold requires |
| PDQ | 80 of 256 | at least 55 | the fingerprint differs more than the threshold requires |
| PSNR | 17.4 dB | no lower than 12 dB | picture quality is fine |
All conditions are met, so the copy got the mark “Unique”. If even one of them had not been met, the copy would have got the mark “Too similar to source”, “Too similar to another copy” or “Low quality”.
Why are there several metrics?
A single metric does not see all changes: SSIM is sensitive to the structure of the frame; fingerprints are robust to small edits and catch copies that look different but remain the same frame; and PSNR does not let you achieve difference at the cost of a degraded picture. That is why a copy passes the check only if all conditions are met at once.
Besides the source, each copy is compared with 100 neighboring copies in the project, with their own thresholds — for example, in “Optimal” the SSIM with a neighboring copy is no higher than 0.92, and the difference by PDQ is at least 35 of 256.
What are ClipClone's thresholds in different presets?
The thresholds are set by the preset. The full table of conditions for all four presets is in the guide “Presets for unique video versions”, and what to do if a copy fails the check is in the guide “How to check that video copies differ”.
The thresholds are the same in all four presets; the presets differ in the strength of the changes, not in the strictness of the check.
| Metric | Threshold in all presets |
|---|---|
| SSIM with the source | no higher than 0.88 |
| pHash and dHash from the source | at least 8 of 64 |
| PDQ from the source | at least 55 of 256 |
| PSNR | no lower than 12 dB |
Frequently asked questions
What SSIM value is considered good for a copy?
There is no universal value: it all depends on how strongly the video needs to be changed. In ClipClone a copy passes the check if its SSIM with the source is no higher than 0.88 — this threshold is the same in all presets.
Why is a high number good for pHash but a low one for SSIM?
SSIM measures similarity: the closer to 1, the more the copy resembles the source. pHash, dHash and PDQ measure the difference between fingerprints: the more mismatched bits, the more the copy differs.
Where can I see the metrics of my copy?
In the Dashboard, in the project: open the copy, and its card will show the check results — the metric values and the thresholds of the preset.
Do good metrics guarantee views in the feed?
No. The metrics show how much a copy differs from the source and from other copies by measurable features. The platform decides what is shown in the feed.
Other guides
- Unique video versions: what they are and how they workWhat making unique versions of a video means, what changes in each version, how the difference between copies is checked and how to make hundreds of versions of one video in the cloud.
- How to make many versions of one videoHow to make many different versions of one video: by hand in a video editor, with software on your computer or in a cloud service. A comparison of the methods and step-by-step instructions.
- Unique video versions for Reels, Shorts, TikTok and VK ClipsHow to prepare different versions of a vertical video for Reels, Shorts, TikTok and VK Clips: the “Strong + vertical” preset, how the copies differ and what the platforms say.
- Presets: Standard, Optimal, Strong, Strong + verticalClipClone presets “Standard”, “Optimal”, “Strong” and “Strong + vertical”: what changes in the copies, which check thresholds each one has and which preset to choose.
- How to check that video copies differ: SSIM, pHash and PDQHow to tell that the copies of a video really differ: what SSIM, pHash, PDQ and PSNR show, which thresholds ClipClone uses and an example of a real check.
- How much unique video versions cost and how output minutes are countedHow the cost of unique video versions is calculated in ClipClone: output minutes, calculation examples, plans from 2,990 ₽ per month, price per minute and the free trial.