Explainer

How Instagram detects a duplicate video, from Meta's own papers

Most pages on this topic guess. Meta has published enough of its own work to say what the matching is trained against, and it is the list every video spinner sells.

What Meta has open-sourced

PDQ is Meta's image hash: a 256 bit fingerprint built from the picture's low frequencies, with a distance threshold for a match. TMK+PDQF and vPDQ extend it to video, one by summarising the whole clip and one by hashing frames and treating them as a bag, which makes trimming and re-timing nearly invisible to it. All three are on GitHub with their thresholds.

What the copy detection is trained against

Meta ran the Video Similarity Challenge in 2022 and 2023 on AugLy, its own augmentation library. The transformations in the training set are horizontal flip, crop, pad, rotate, speed change, time cropping, colour jitter, brightness, greyscale, noise, blur, pixelation, encoding quality, text and emoji overlays, stacking, screenshots and perspective shake. In plain words: every edit a "make my video unique" tool offers is on the list the detector was built to see through. A learned detector in that challenge lost about seven percent of its accuracy under speed manipulation, which is a small, known weakness.

What Instagram has said on the record

• April 2024 and April 2026: accounts that repeatedly repost other people's content are taken out of recommendations. A watermark, a speed change or a screenshot with credit still counts as unoriginal. Unique text, creative edits and a voice-over count as original.
• Since 2025, creators report a notice on a trial reel that resembles one already on the same account, saying reach will be suppressed. So the comparison also runs within your own account.

What that means for your own reel

Reposting your own footage is outside the aggregator rule, and the trial-reel similarity check still applies to it. The technical edits alone are a coin flip. The changes Instagram itself names as original are a new line of text, a creative edit and a voice-over, which is why a re-cut of your own reel should start with a new hook line and add one more visible change on top.

What nobody has measured

No threshold is published. The "seventy percent similarity" figure and the "first six seconds are scanned" line exist only on marketing blogs. Metadata and file hash changes do nothing against a perceptual matcher by construction. Anybody who promises a repost that Instagram cannot see is guessing, and Meta's own training list says the guess is a poor one.