What is a video fingerprint?

Two families of hash

People use “hash” for two different things, and the confusion causes most bad advice in this niche.

	Cryptographic hash (MD5, SHA-256)	Perceptual hash (pHash, dHash)
Input	Every byte of the file	The visual content of a frame
Changes when	One byte changes	The picture changes
Survives re-encoding	No — completely different	Yes — stays similar
Used for	File integrity, exact duplicates	Content matching, recommendations, duplicate detection

A re-export changes the cryptographic hash and nothing else. That is why “the MD5 is different” is not an argument that a video is unique.

How a perceptual hash works

Most perceptual hashes follow the same recipe: reduce the image, simplify it, then encode relative differences as bits.

1. Downscale. A 4K frame becomes a tiny grid — commonly 9×8 or 32×32 pixels.
2. Remove colour. Convert to grayscale; colour is easy to change and noisy for matching.
3. Compare neighbours. dHash records whether each pixel is brighter than the one to its right, one bit per comparison.
4. Store the bits. Sixty-four bits for an 8×8 dHash — small enough to compare millions of times.

Two similar frames produce hashes that differ in only a few bits. The number of differing bits is the Hamming distance: 0 means identical, higher means more different.

From frames to a video fingerprint

A video is not one image, so production systems do more than hash frame one:

• Sample keyframes. Hash representative frames across the timeline.
• Compare sequences. The order and timing of similar frames matter, not just a single match.
• Scene detection. Split on cuts and fingerprint each scene, which makes partial reuse detectable.
• Embeddings. Some systems use learned vectors rather than hashes; the principle is the same — a compact representation that survives transformations.

Public platforms do not publish which method they use. Everything above is the class of technique, not a claim about a specific company's pipeline.

The audio fingerprint

Speech and music are summarised differently: the system looks at spectral peaks over time and stores their pattern. This is robust to compression, background noise and moderate volume changes, which is why audio is often the strictest layer. A video re-encode leaves it almost untouched.

That is also why removing or changing the audio is not a free win: removing it changes the experience, and small pitch/EQ/timing changes are the only way to alter the fingerprint without ruining the sound.

What a fingerprint is not

• Not a watermark. Watermarks are visible marks; fingerprints are derived numbers.
• Not metadata. Encoder and device tags are a separate signal, not the fingerprint.
• Not a guarantee. A fingerprint comparison is probabilistic; thresholds are set by the platform, not by math.
• Not ownership. Two identical files can belong to different rights holders; the hash says nothing about rights.

Measuring distance yourself

Compute a dHash for the source and for each output, then count differing bits. In RepurposeThis the uniqueness report does exactly this and reports a median for the batch; the Fingerprint Lab goes further and ranks individual effects.

What the number means depends on the hash size. A 64-bit dHash with a distance of 10 is meaningfully different; a distance of 1 is not. There is no universal “pass” threshold, and any vendor quoting one should explain where it came from.

Short answers

---

Next: nine changes ranked by impact, or the glossary.