Add your own categories to photo keywords, automatically

Auto-tagging in photo tools and DAMs describes what's in the picture. It can't know that a photo is PN-10423 rather than PN-10424, or water damage under your claims policy rather than staining. Those are your decisions, and only your past photos record them.

Photo libraries, DAMs, inspection apps and cameras write the category into the photo's own metadata, in one of three places:

• Keywords (IPTC Keywords, XMP dc:subject, or Windows Tags): the usual place.
• Title: often a product name or part number.
• Description (caption): sometimes a code with a sentence around it.

A photo can carry several keywords: parcel, warehouse 4, category:Crushed corner. Then a pattern says which one is the category, for example ^category:(.+)$.

With StayCharted AI Model Trainer:

1. New model → Pictures. Upload a ZIP of photos, a spreadsheet of photo links, or connect your storage (Amazon S3 or S3-compatible).
2. Where is each picture's category? Choose In the picture's own metadata, and the field: Keywords, Title or Description. The screen shows how many photos have each one ("Keywords (on 34)").
3. Preview the categories. You see the categories, how many photos each has, and which photos were left out and why: no value in that field, or several values and no pattern to pick one.
4. Check your pictures, then train. The model is tested on photos it never trained on, and the report shows how often it's right per category.

The model learns from the picture itself, not its keywords. The keyword is only the label. A new photo with no metadata at all is answered the same way.

Categorize new pictures, and give it a ZIP of new photos. Under What you get back, tick Add the category to each picture's metadata as its Keywords (or Title, or Description). If the model learned from a metadata field, that field is already chosen.

You get:

• The photos back, as a ZIP, under the same names, each with its category written in.
• Photos the model is unsure of go in a _needs-review/ folder in the ZIP with nothing written, and into the Review Queue. A guess never lands in your library as if it were known.
• The spreadsheet of each photo's name, category and confidence, as always.

If the model learned the category through a pattern, it's written the same way (category:Torn tape), so the next training run reads it back.

Want only the spreadsheet? Leave the box unticked. No photos are made.

From your storage: with a folder for sorted copies, each copy gets its category written in, and the original is never touched. See Connect your storage.

On your side: if the photos are already in your library, have it re-read the files, for example Metadata → Read Metadata from Files in Lightroom Classic.

On ours: the files you upload are kept as you uploaded them, for the applicable retention window. Original training ZIPs are deleted after 90 days unless deleted sooner; fill uploads are deleted when the fill finishes, or after seven days if never run. See the Privacy Policy. AMT's own small copy of each picture, the one the model reads, is made from the pixels alone. Location and other camera details are never read into your model.

Be consistent first. If half your library says Water damage and half water-damaged, that's two categories. Merge them in your library first, or change them on AMT's Check your pictures screen.

• Photos with several keywords and no pattern are left out of training, and listed. Add a pattern, or a clear category keyword.
• Old tools that read only IPTC (IIM) keywords may not show the new keyword, because it's written in XMP, the current standard. Check one photo in your tool before a big batch.
• Unsure photos aren't keyworded. That's on purpose. Check them in the Review Queue.