AI Image Tagging for Fashion Catalogs

A fashion catalog does not stay a folder for long. One season of flat-lays, packshots and on-model shots, plus every reference image and every generated variation, and the library holds thousands of files that nobody named consistently. At that point image tagging stops being housekeeping. It becomes the layer that lets every other step of the pipeline find the right image without a person scrolling for it.

This guide covers what image tagging is, what AI can tag on its own and what it can't, how to structure tags for a fashion catalog, and how it works in On-Model, where tagging is part of the production system rather than a feature bolted onto a gallery.

What is image tagging?

Image tagging is attaching short labels to an image so it can be found, filtered and acted on by meaning rather than by filename. A tag can describe what is in the image (a denim jacket, a model, a white background) or record something people decided about it (approved, SS26, for Zalando).

AI image tagging, also called automatic image tagging, is the first kind done by a vision model: the image is analysed when it arrives and labels are written without anyone typing them. Manual tagging is the second kind, and no model can do it for you, because "approved by the brand team" is not visible in the pixels.

Two kinds of tags, two owners

Most tagging problems come from treating these as one thing. They have different sources, different vocabularies, and they answer different questions.

Automatic image tagging is very good at the left column and useless for the right one. A good system does the left column for you and makes the right column cheap to maintain. A bad one asks people to type "shirt" four thousand times, or tries to guess approval states from pixels.

What AI image tagging gets right in fashion, and where it slips

Generic image-recognition models were trained on everyday photos, so they know "shirt" and "shoe" but describe fashion in a vocabulary that is too coarse for a catalog. Four things separate a tagger that works on fashion from one that doesn't:

• Garments and details are kept apart. "Jacket" is a product; "collar", "zipper" and "neckline" are parts of it. When they land in one list, a search for jackets returns every shirt with a collar.
• Colour is named the way merchandisers name it. "Dark green" is useful, "green" is barely useful, and a hex code is not a search term anyone types.
• Material and construction are visible and worth tagging. Fleece, knit, denim, leather and quilted are what a buyer searches for, and a vision model can see them.
• Presentation matters as much as content. A shirt on a model, on a hanger, or laid flat are three different assets for production, even though a generic tagger calls all three "shirt".

Where every tagger slips: small text (care labels, prints with lettering), near-identical colourways in one batch, and anything that depends on context outside the frame. Use descriptive tags to find candidates, never as the final word on what ships.

How to structure tags for a fashion catalog

Keep the operational vocabulary short and shared. Three lists cover most teams:

List	Example tags	Why it pays
Review state	Ready, Needs retouch, Approved, Rejected	Replaces the most chat messages; the first thing anyone filters by
Season or campaign	SS26, Winter27, Holiday drop	Cuts a year-round library back to the shoot you're working on
Destination	PDP, Amazon, Zalando, Paid social	Channels have different specs; one tag per channel avoids re-checking

Two rules make this hold up over a year. Agree on names before tagging (SS26 and ss-26 are two different tags to a filter), and keep descriptive words out of operational lists, because the AI already covers them.

How image tagging works in On-Model

On-Model treats tagging as infrastructure: three layers, each written by a different owner, all on the same images, all usable as filters and through the API.

1. Automatic labels on every upload

Every image that lands in the library is analysed automatically within seconds. The system detects the clothing and accessories in it (shirt, pants, dress, sneakers, sunglasses, bag) and person attributes (adult, child, female, male), and shows them on the asset as read-only Detected by AI labels. There is nothing to enable and nothing to type.

2. A description and keywords the first time an image is used

When an image goes into a tool for the first time (Flat-to-Model, Create Packshot, Model Swap, Garment Recolor or Detail Repair), On-Model looks at it more closely: it writes a one-sentence description and a set of keywords for colour, material, parts and styling. A dark green fleece gets jacket as its garment and dark green, fleece, collar, zipper as keywords, kept in separate rows for the reason above.

This understanding is stored on the asset once and reused everywhere after that: by every later job, by every tool, and by the library search. The search bar on the Assets page matches detected garments, keywords and descriptions, so typing fleece or zipper finds the images that contain them, whatever the files were called when they were uploaded.

3. Your team's tags, shared and actionable

The operational layer is Tags: a name and a colour, applied to generated results, uploaded assets and saved identities from one shared library. They show as coloured dots on thumbnails, and every gallery and list screen can filter by them. They are visible to everyone in your team with no share step, free on every plan, and cost no credits. The full walk-through is in Tag Your Images. Find Them Later.

What makes this infrastructure rather than labelling: tags drive production, not just retrieval. Bulk Create starts one Flat-to-Model or Create Packshot job per tag, so tagging a folder of products by SKU is how a whole collection gets processed in one run. Smart tags can set those tags at upload from folder names or filename prefixes, and every output inherits its product's tag, so the results come back already sorted.

Tagging through the API

When On-Model sits inside a larger pipeline, tags travel with the images. Files can be uploaded already tagged, tags can be assigned to assets, identities and job outputs, and the assets, identities, job-results and gallery endpoints all accept tag_ids to return only what carries them. Your DAM can mark what it ingested and your review tool can write approvals back, without anyone opening the app. Reference: Tags API docs.

Choosing image tagging software: a checklist

Whether you evaluate a DAM, a standalone tagging API or a production platform, the same questions separate useful from decorative:

1. Does automatic tagging use a fashion vocabulary? Ask for garments, colours, materials and parts as separate fields, not one bag of words.
2. Are detected tags searchable? Labels you can only look at do not save time. Labels you can search do.
3. Are your own tags shared with the team by default? A tag only one person can see is a bookmark, not a workflow.
4. Can tags be read and written by API? Otherwise every system around it needs a person in the loop.
5. Do tags trigger anything? The difference between a tagged library and a pipeline is whether a tag can start work, not just filter it.
6. What does tagging cost per image? Tagging that is billed per image gets switched off on large libraries, which defeats the point.

Common questions

What is AI image tagging? AI image tagging is the automatic labelling of images by a vision model: each image is analysed when it is uploaded, and labels for what it contains (garments, colours, materials, people) are written without manual input. It makes a large library searchable by content instead of by filename.

Is automatic image tagging accurate enough for a fashion catalog? For finding images, yes: garments, colours, materials and presentation are reliably visible. It is not reliable for small printed text, near-identical colourways, or anything decided outside the image, such as approval state or channel. Use automatic tags for search and human tags for decisions.

What is the difference between manual and automatic image tagging? Automatic tagging describes what is in the image and is done by AI. Manual tagging records what your team decided about it, such as Approved or SS26. A catalog needs both, and they work best kept as separate layers.

Can I search my On-Model library by what is in the images? Yes. The Assets search matches the garments, keywords and descriptions On-Model detects, so glasses or fleece finds images containing them. Your own tags are applied through the Tags filter on each screen.

Does tagging cost credits in On-Model? No. Detection, keywords and your own tags are free on every plan, and tagging never changes how a job is processed.