I started Schmarchive because my photo library has been steadily growing by me dumping stuff from every phone, camera, backup, and backups of backups on a drive. And I kept procrastinating and never organized it. There is a particular despair instilled by looking at a filename like IMG_20250703_143012.jpg. It tells me when the camera thinks the picture was taken, but not where I was, what I was looking at, or whether I already have four almost identical copies somewhere else. I wanted a small tool that could weed out unwanted, redundant and pointless photos.

Enter Schmarchive: a command-line archive organizer that non-destructively does that. It can rename pictures from metadata stored as EXIF dates and GPS, reverse-geocode locations, and use that to scaffold a folder structure, find duplicate or near- duplicate images. A local vision model can then help decide which repeated shot is best, sort photos into broad categories, or find every image containing something like a white cat.

The unglamorous bits are the interesting bits

The first version was mostly metadata and file operations. That is still the heart of it. Before asking a model anything, Schmarchive extracts file hash for comparing exact copies and a perceptual hash for images that look similar. Only the candidate groups are sent to the vision endpoint. The model confirms the group and picks the sharpest, best-exposed, best-composed image. The remaining files are renamed with a duplicate suffix before they can be moved aside.

That intermediate step matters. File management is one of those areas where a confident wrong answer is worse than no answer. I would rather see photo__near_dupe.jpg and decide what to do than discover that an assistant quietly deleted a memory.

Geotagging

Geotagging has a similar practical compromise. Calling a free reverse-geocoding API (Nominatim) once per image would be wasteful and impolite, so photos with nearby coordinates are first clustered by a configurable distance and one result is cached for the group. The file gets a name such as cordoba-2025-03-24.jpg, while the geocoding work is kept in a local locations.csv cache for later runs.

Why local AI?

I wanted the ability to pluck photos by subject, e.g. “a white cat”, which prompts the vision model to return Detected / Not detected for that subject. If no subject is provided, it can also do it by a preset list of generic categories e.g. nature, cityscape, nightshot etc. which can classify photos without supervision.

Simple renaming, copying, moving, as well as any other deterministic action is done programmatically.

Other goodies

Metadata is inconsistently present, model can be wrong, and my archive is full of old photos. But it solves a real problem for me: turning an overwhelming number of files into something I can enjoy.

Here are the full list of commands you can run:

  1. Geotag files — rename photos by location and date.
  2. Date-tag files — rename photos using capture dates.
  3. Deduplicate — find exact and near-duplicate images with hashes and AI verification.
  4. Move duplicates — move marked duplicates into a separate folder.
  5. Normalize filenames — replace accented/non-ASCII characters.
  6. Organize photos — sort by year/month/location or AI-detected subject.
  7. Identify landmarks — rename photos using GPS-based landmark information.
  8. Pluck images by subject — find and move images matching a description.
  9. Configure settings — change folders, model, API, thresholds, and related options.
  10. Process videos — geotag and organize videos by date and location.