- SauceMatch
- How it works
How our reverse video search works
What happens to your file, how the frames are chosen, why a Match is proof and a Similar result is not, and what is in the index.
Why Google and TinEye miss adult clips
Google Lens and Google Images filter explicit results, and general reverse image engines index the open web: pages and the pictures on them. The frames of a video are not on a page. What tube sites do publish are preview stills: the thumbnails in their listings and the frames that play when you hover a video. This index is built from those previews, 46,348,537 of them from 5,218,447 videos, so it can find the video a still came from.
1. Your file, on your device
Your file never leaves your device: the frames are chosen in your browser. On most computers they are fingerprinted there too and the fingerprints are sent; on phones and slower devices, and when that does not work in your browser, the chosen frames are sent and our server does the same steps. For a video, the page:
- Scans it at the preview positions. Sites cut their previews at even points of the whole video: 16 steps (Pornhub, Tube8, RedTube), 17 (YouPorn), 15 (Eporner), 9 (XVideos, XNXX), 20 or 60 elsewhere. The scan visits the same kinds of points, centre and start of each step, plus an even sweep in between: up to 160 positions, whatever the length.
- Looks at each position. Brightness, contrast and sharpness are measured on a small copy. Black frames, fades and motion blur are dropped; near-duplicates are merged.
- Crops what is not the video. Black bars and the blurred copy that phone apps put around a horizontal video are found across many frames and cut away.
- Keeps the best 48. At least one per stretch of the timeline, then the sharpest and most distinct, with a bonus for the preview positions.
A GIF is decoded frame by frame and treated the same way; a picture is searched as it is. A video your browser cannot play (some MKV, AVI or HEVC files) cannot be searched here: convert it to MP4, or search with a screenshot from it.
2. The search
Each frame becomes a 512-number fingerprint from a copy-detection network (SSCD), trained to give the same fingerprint to a picture and its edited copies: crops, scaling, colour shifts, text overlays, recompression. On most computers this happens in your browser, which downloads the network once and keeps it for the next visit; on phones and slower devices, and when it cannot happen in your browser, it happens on our server. The index answers with the nearest stored frames and the videos they come from; the fingerprints are not stored. Frames that appear in more than five videos (logos, intros, placeholders) are ignored.
3. The proof
Fingerprints find lookalikes; they cannot prove identity. The same room filmed on another day, or the same performer in a similar pose, can look close. So every promising candidate is checked geometrically, again in your browser (or on our server, when it did the search): its stored picture is downloaded, keypoints are found in both frames and paired, and the pairs must agree on one transform (a homography) that maps your frame onto the indexed one. Only then is it a Match, with a confidence from 0 to 100.
The gates were calibrated on held-out tests: re-uploads that were really the same video never scored below 28 agreeing keypoints, while the same set in a different video stayed below the gates. A video that does not pass but stands out from the rest of the index across the scenes of your clip is listed as Similar: a lead worth a look, unverified, never shown as an answer. A picture that several indexed videos share, such as a studio's intro or outro card, is listed separately: it usually says who made the clip, not which video it is, so it does not count towards Similar.
4. A closer look when it is close
When a video comes up as Similar, the page samples your clip again a few seconds either side of the moment that found it and searches those frames (never around a shared intro or outro picture): the index holds only a few stills per video (about 20 at most), and the exact one may sit between two of the first pass. When nothing matches at all, the best frames are also searched mirrored, since reposts are often flipped.
What is in the index
46,348,537 frames from 5,218,447 videos, collected from public preview images. The search index is rebuilt with the new ones twice a day.
- reddit.com
- xvideos.com
- eporner.com
- thisvid.com
- missav.ws
- porntn.com
- redtube.com
- pornhub.com
- namethatporn.com
- femdomhd.org
- erome.com
- mixfemdomcc.com
- hotpic.cc
- clips4sale.com
- hypnotube.com
- menareslaves.com
The same search runs behind every page of this site, whichever way you start it: a reverse video search with a clip, a reverse image search with a screenshot, or a reverse GIF search.
Anything can be removed on request: see removal requests.