A deduplication tool can be wrong in two ways. It can miss a duplicate: you keep one copy too many, which can be fixed. It can also invent one: two different photographs are presented as the same, you discard one, and it is lost. Obscura Trace is tuned so that the second error is as rare as possible. When in doubt, it keeps.
An exact duplicate is a fact
Two files are exact duplicates when their content is identical byte for byte. The app establishes this with their full SHA-256 checksum. To avoid reading everything, it works in three steps:
- only files of the same size can be identical;
- among them, a partial checksum, made from a few portions of the file, quickly rules out those that differ;
- only the remaining candidates are read in full to compute the SHA-256.
Only the full SHA-256 makes an exact duplicate. Name, date, or size are never enough. A candidate whose full checksum has not been computed yet is shown as “Doublon exact probable, hash complet requis” (probable exact duplicate, full hash required), never as exact.
An exact duplicate carries an established confidence: “Contenu identique · 100 % · SHA-256 complet vérifié” (identical content, full SHA-256 verified), with a checked seal and without the word “estimé” (estimated).
A probable duplicate is an estimate
A re-saved JPEG, a resized version, or a black-and-white conversion is not identical byte for byte. Yet it is the same photograph. To match them, the app computes two visual fingerprints for each image, dHash and pHash. It matches two images only if both fingerprints are close: at most 10 bits of difference out of 64.
This double criterion is deliberately strict. A probable duplicate carries an estimated confidence, always followed by the word “estimé”:
| Qualifier | Threshold |
|---|---|
| “Même négatif” (same negative) | 95% or more |
| “Probablement la même photographie” (probably the same photograph) | 80% or more |
| “Possiblement liées” (possibly related) | below |
An estimate is never shown as a fact. It is confirmed in the Review Center.
What visual fingerprints do not see
The measured limits are published, not hidden:
- cropping is not recognized;
- rotation is not recognized;
- strong contrast changes can make a match go missing.
Several scans of the same negative, cropped differently, will therefore generally not be grouped in the current release. That is the price of cautious thresholds: a missed match can be caught by eye, an invented one costs a photo.
The master is a recommendation, not a probability
In each group, the app recommends the copy to keep first, the master. It applies a seven-criterion rule scored out of 100: resolution, bit depth, format, metadata, color profile, size, capture date. It explains its choice criterion by criterion, naming the files.
That score is not a probability. It is the result of a rule. Its level tells you whether to look closely: “Recommandation nette” (clear recommendation), “Recommandation à confirmer, écart faible” (to confirm, small gap), “Égalité stricte, master à confirmer” (strict tie, master to confirm).
You decide, nothing moves
In the Review Center, each decision has its button and its shortcut: “Choisir comme master” (choose as master), “Conserver” (keep, ⇧⌘K), “Marquer comme version” (mark as version, ⇧⌘V), “Marquer comme doublon” (mark as duplicate, ⇧⌘D). No decision touches a file. Decisions are recorded in the catalog, durable, and undone with “Rouvrir le groupe” (reopen group).
They are then used by the transfer: duplicates you set aside are not copied, and stay in the source. Nothing puts them in the Trash.
The detail is in the Understanding the concepts and Review Center chapters of the guide.