What we tested, and what it can and cannot prove
Instagram's own matching system is private, so nobody outside Meta can test against it directly. As a public stand-in, we used Meta's open-source matching tools from its ThreatExchange project: the TMK+PDQF video hash and the PDQ frame hash. These are real matching tools, but they are not proof of what Instagram runs internally.
The video hash gives a similarity score where 1.00 is an exact match. Following the tools' own guidance, we treated 0.70 as the match line: at or above it, the matcher calls two files the same video. Below it, the copy passes.
The controls
First, two sanity checks to confirm the matcher works as expected:
| Test file | Score | Result |
|---|---|---|
| Plain re-upload of the same file | 1.00 | Matched, as expected |
| Mirror-only copy | -0.22 | Passed, but scored 1.00 once we un-mirrored it |
The mirror result matters: a mirrored copy only passes because the matcher does not flip frames back on its own. Once we un-mirrored the copy, it matched perfectly. A mirror alone is a weak edit.
Single edits vs the video hash
We applied one edit at a time and scored each copy against the original:
| Edit | Score | Result |
|---|---|---|
| Hue shift | 0.999 | Matched (useless) |
| Gamma / contrast | 0.999 | Matched (useless) |
| Speed +6% | 0.999 | Matched (useless) |
| Rotate 2 degrees | 0.91 | Matched |
| Rotate 3 degrees | 0.81 | Matched |
| 10% crop on one side, stretched | 0.82 | Matched |
| Zoom 10% + rotate 1.2 degrees + shift | 0.71 | Borderline |
| Off-center crop 9% to a corner + saturation | 0.60 | Passed |
| Off-center crop 12% | 0.35 | Passed |
| Off-center crop 15% | 0.09 | Passed |
The pattern is clear. Color, contrast and speed changes barely move the score. Rotation and simple crops help but stay above the line. Off-center reframing is the edit that actually separates a copy from its original, and 12 to 15% off-center passes with room to spare.
Frame hash (PDQ)
The PDQ frame hash compares individual frames rather than the whole video. It was easy to beat: no copy matched, with distances of 48 to 104 against a match line of about 31. Whole-video matching is the harder test, which is why we weight it more.
Audio fingerprint
On our test clips, a 3 to 4% pitch shift with speed unchanged moved copies past the audio match line. A 2% shift was not enough. Tempo changes or EQ alone did nothing. If a tool only adjusts brightness and never touches the audio, the soundtrack still matches.
AI copy detectors
We also tried Meta's open-source SSCD model, which is built to recognize the same content even after heavy edits. It recognized every reasonable edit as the same footage, even a 30% zoom. Color changes barely registered. If a platform runs a model like this, file edits alone will not hide that two posts share the same footage, and no honest tool should claim otherwise.
The TikTok field test
We posted the same clip across 20 TikTok accounts. It was fine on 17 and flagged on 3. Even an exact repost posted fine on some accounts. Same file, different outcomes.
The takeaway: on TikTok, account health and adding real new input (new hook text, a new sound) matter more than file edits. File edits still matter for giving each account its own version, but they are not the whole story on that platform.
What we changed because of it
In Auto, every Fresh Copies copy is reframed 12 to 15% off-center, the edit range that passed with the most margin, plus a 3 to 4% audio pitch shift. Small brightness, color, contrast and grain changes still stack on top, and every copy gets a uniqueness score. In Auto, copies that score too close to the original are remade automatically.
No test can guarantee how a platform will treat a post. What we can do is measure our copies against public tools, publish the numbers and keep the bar where the data says it should be. Check your own copies with the free copy checker, or read why TikTok flags some copies for the account side of the picture.
Common questions
Does this prove what Instagram uses?
No. Instagram's matching system is private. We tested against Meta's open-source TMK+PDQF and PDQ tools as a public stand-in, which is the closest honest benchmark available.
Which single edit works best?
In our tests, off-center reframing. A 12 to 15% off-center crop scored 0.35 to 0.09 against a 0.70 match line, while color, contrast and speed edits scored 0.999 and changed nothing.
Is mirroring a video enough?
No. A mirror-only copy scored -0.22, but matched 1.00 as soon as we un-mirrored it. Treat mirroring as a small extra, not a real edit.
Can any edit guarantee a post won't be flagged?
No. No tool can guarantee how a platform treats a post. Our TikTok field test showed the same clip flagged on 3 of 20 accounts, so account health and new creative input matter too.
Try it on your own video
Free trial: 20 copies. No card. You can also test copies you already have with the free copy checker. It runs on your device, so your files never leave it.