videre stats

videre stats                           # library totals and inventory
videre stats --json                    # print one JSON object instead
videre stats --mismatched             # list every filename/content mismatch
videre --library ~/Photos stats        # select a different library

Pipeline health, staleness, watch liveness and what to run next live in videre status, not here: stats is what is in the library, status is what state the pipeline is in.

Reading the output

Library: 70601 file(s) (398.9 GB), 57142 photo(s), 13459 video(s)
Duplicates: 0 group(s), 0 file(s), 0 B wasted
Faces: 58555 detected, 86 people named
Marks: 312 rated, 40 picked, 12 labelled, 128 liked

Embeddings:
  google/siglip2-base-patch16-224           70588   768-dim   128.3 MB
  google/siglip2-so400m-patch14-384         70587  1152-dim   189.6 MB
  google/siglip2-base-patch16-384           70588   768-dim   128.3 MB

By type:
  mov      video/quicktime             12564   165.4 GB
  jpeg     image/jpeg                  43524   142.4 GB
  mp4      video/mp4                     643    62.0 GB
  heic     image/heic                  10395    20.8 GB
  png      image/png                    3211     6.5 GB
  mov      video/mp4                     251   909.5 MB
  mp4      video/quicktime                 1   462.5 MB
  dng      image/tiff                      9   433.9 MB
  gif      image/gif                       2     4.8 MB
  png      image/jpeg                      1    43.7 KB

Mismatched files: 1
  A mismatch means the filename and last scanned content disagree; verify before renaming.
  image/jpeg  /Photos/sunset.png

Disk use:
  embeddings           446.3 MB
  database             161.8 MB
  thumbnails            84.1 MB  (rebuildable)
  place names           11.4 MB  (rebuildable)
  database journal      32.0 KB  (rebuildable)
  total                703.7 MB  (95.6 MB of it rebuildable)

Real output from a 70,000-file library, so the awkward rows are the interesting ones.

By type groups the library by file extension, with the mime type beside it and the largest first. These are the values --ext and --mime take, so this is where to look before scoping a run.

By type preserves both values, even when they differ. The .mov files detected as MP4 video and the .mp4 detected as QuickTime are normal container brand differences, not suspicious filenames. The .png detected as JPEG is a filename/content mismatch. Keeping extension and detected MIME separate means --ext mov and --mime video/quicktime still select genuinely different sets.

Mismatched files counts paths whose known extension conflicts with their known detected MIME, then shows the first 10 paths in path order. Run videre stats --mismatched for the full list; ordinary stats says how many more paths were omitted. A zero count is shown explicitly. The report uses the last scan’s stored MIME, not a fresh read of the media. It does not imply videre damaged the file or that the file can be decoded; inspect it before renaming anything. Files with unknown MIME are not reported as mismatches. Printable characters in paths, including Unicode combining marks, are shown as-is. Control characters and invisible direction or zero-width characters are escaped, so a filename can neither alter the terminal nor display reordered.

Disk use is what videre itself stores, not your photos, largest first. Entries marked (rebuildable) cost only time to recreate: thumbnails are re-converted on demand, place names ship with videre, and the database journal is transient. The database and embeddings are not rebuildable - an embedding run takes hours, and the database cannot be recreated without rescanning.

A logs row counts .videre/logs/, every command’s log and its rotated copies. It is marked rebuildable because deleting it loses only history, never library data; how much it keeps is set by the log settings in logging.

Note the proportions above: 703.7 MB of videre data describing 398.9 GB of photos, and only 95.6 MB of it disposable. Embeddings dominate because that library has three models prepared.

Embeddings are counted per library

They live under this library’s own .videre/embeddings/, so stats reports only the ones belonging to the library you asked about. Another library’s embeddings live under its own root and never appear here. See keeping libraries separate.

Duplicates counts exact copies only: the same pixels or media data, even when their metadata differs. “Wasted” is what you would reclaim by deleting all but one of each group, which is what videre dedupe proposes.

Faces detected counts individual faces, not photos, so one group shot contributes several - 58,555 faces across 57,142 photos above. People named stays at 0 until you assign names in gallery; detection alone never produces a name, however many faces it finds.

Embeddings lists one line per model you have prepared. The dimensions come from the stored data rather than a hardcoded table, so an unfamiliar model still reports honestly - which is why the 1152-dimension so400m line above needs no special-casing to appear correctly.

The counts differ slightly between models (70588, 70587, 70588) because each run skipped whatever it could not decode at the time. A count below the library total is normal, not a sign of a failed run.

JSON output

--json prints one object: schema_version, library (the totals, faces, marks and embeddings above), by_type (every type, not only the top 12 the text shows, each with ext, mime, files and bytes), mismatches (the exact count, a files array of {path, ext, mime}, and truncated), and disk_use (each row with label, path, bytes, files and rebuildable). The JSON mismatch list also defaults to 10 paths; combine --json with --mismatched for all paths. It remains one JSON object even when the count is zero.

Numbers that look wrong but are not

Photos plus videos may not equal the total. In the example, 151 + 52 = 203 against 204 files. The missing one is a file whose type could not be identified from its bytes, so it counts toward the total but is neither a photo nor a video.

The embedding count is lower than the file count, and should be. Vectors are keyed by content, so duplicate copies share one, and .dng files are never embedded at all. A gap here is normal; a gap of thousands means videre embed has more to do.

duration=0ms is not an error. Commands reading an already-warm database genuinely finish in under a millisecond.

A command can show success and still have exited nonzero. Per-item problems, a few unreadable files or one corrupt image, do not fail a run. Only an unhandled error does. fix-dates and faces both return a count of problems rather than failing outright.

Pipeline health moved to status

Per-command last-run lines, the status table (success, failed, interrupted, crashed), watch liveness, and the --check exit code for cron all live in videre status now. Keeping them in both places would let the two surfaces disagree; one home, and it is status.