STAYCHARTED BLOG · MACHINE LEARNING · PRODUCT IMAGES
Machine learning for product image classification: sort catalog photos into your own categories
Machine learning for product image classification: 91.1% on 40 catalog photo categories, 83.5% with 5 photos each, and how to sort or tag your own photos.
What a product image classification model does
A supervised model learns from pairs: a photo and the category your team gave it. After training, it looks at a new photo and returns:
- a category from your list
- a confidence from 0 to 1
- a needs-review flag when the confidence is below the cutoff you set
The model learns from the image itself, not from the file name or folder name, so a new photo with no name or metadata is classified the same way. Answers below the cutoff go to the Review Queue, and reviewers' answers train the next version. For catalog image categorization, that means new supplier photos, returns photos or user uploads can be sorted into the same taxonomy your team already uses.
Product image classification accuracy: what the catalog benchmark shows
We compared three open image models as the reading step for the AI Image Classifier Model and published the results in why AMT's image model uses SigLIP 2. The product test used a public set of fashion catalog images, keeping the 40 product types with at least 60 pictures each: 7,966 pictures to train on and 1,989 to test on. Each model was also trained with only 5, 10, 20 or 50 photos per category.
Accuracy on the 1,989 held-out catalog product photos:
| Training photos per category | SigLIP 2 | DINOv2 small | DINOv2 base |
|---|---|---|---|
| 5 | 83.5% | 73.2% | 77.7% |
| 10 | 85.9% | 79.0% | 81.9% |
| 20 | 87.5% | 81.0% | 83.9% |
| 50 | 89.8% | 84.7% | 86.4% |
| All training data | 91.1% | 88.3% | 88.8% |
SigLIP 2 led on product photos at every training size, and the gap was widest with few examples. At 5 photos per category it was 5.8 points ahead of DINOv2 base and 10.3 points ahead of DINOv2 small. That matters for product photo classification, because new or seasonal categories often start with a handful of photos.
The page offers one interpretation, which it labels as not directly tested: SigLIP 2 learned from pictures paired with descriptions, "so it has absorbed what things are called." Product categories are mostly about what a thing is, which fits that kind of training.
A fine-grained comparison: 37 pet breeds
The same page tested 37 cat and dog breeds (3,680 training and 1,850 test images). Here the result went the other way:
| Training photos per category | SigLIP 2 | DINOv2 small | DINOv2 base |
|---|---|---|---|
| 5 | 85.5% | 87.3% | 90.2% |
| 50 | 94.2% | 94.2% | 95.7% |
| All training data | 95.2% | 95.2% | 96.0% |
DINOv2 base led on breeds at every size. Breeds are a fine-grained visual task (same kind of animal, small differences in shape and coat), and a model trained on images alone did slightly better. We chose SigLIP 2 for catalog-style sorting, and the page says plainly that other business imagery may favor a different model.
Speed on an ordinary CPU
SigLIP 2 "ran at about 20 pictures a second on two laptop CPU threads, including decoding and resizing." DINOv2 small was about 2.4 times faster, but gave up 3 to 10 points on product photos. For most catalogs, the accuracy was worth more than the speed.
Where a fine-tuned image model fits, and where it lost
AMT also offers the Dedicated AI Image Model, which fine-tunes the image model itself (Business plan with the Vision add-on). The benchmarks page lists no Dedicated AI Image Model result for catalog product photos, so we do not claim one.
On other photo tasks the results are mixed, and the page shows both directions. On car damage photos (6 categories, 552 photos) the Dedicated AI Image Model scored 86.4% against 76.6%. On 100 look-alike part numbers with 20 training photos each, the AI Image Classifier Model scored 87.8% and the Dedicated AI Image Model 75.3%. The page suggests few examples per part as a possible reason. Start with the AI Image Classifier Model, and try the Dedicated AI Image Model when you have many photos per category and the held-out test shows it helps.
How to start: folders, storage or keywords
You probably already have sorted photos somewhere. AMT can learn from any of these, and needs at least 10 photos per category (20 or more is better).
Folders of sorted photos
Put each category's photos in its own folder and upload the folders as a ZIP: each folder name becomes a category. If the ZIP is larger than your plan allows, upload a spreadsheet with a link to each photo and its category instead. See train an image classifier without code.
S3-compatible storage
Connect Amazon S3, Cloudflare R2, Backblaze B2, Wasabi, DigitalOcean Spaces, Google Cloud Storage (with an HMAC key) or another S3-compatible endpoint. In Settings, you list the folders AMT may read and, optionally, the folders it may copy into. When you create a model, you choose where each photo's category comes from: a folder name at a level you pick, the file name or a metadata field. Azure Blob Storage is not supported yet.
When you fill new photos from a folder, you get:
- a spreadsheet of each photo's category and confidence
- unsure photos in the Review Queue
- optionally, sorted copies in
<folder>/<category>/, with unsure ones in_needs-review/ - optionally, tags on those copies (category, confidence, review and model) and the category written into each copy's metadata
Your originals are never moved, re-encoded, tagged or deleted. Full details are in classify images in an S3 bucket.
AI image tagging for ecommerce photo libraries
If your team keeps categories as photo keywords, AMT can read them from IPTC Keywords, XMP or Windows Tags, or from the Title or Description field. New photos come back with the category written into their metadata as a keyword, title or description, in the same form so the next training run can read it. Image data stays byte for byte identical, JPEG and PNG are written, and Lightroom and Bridge read the result. Unsure photos are left without a keyword and go to the Review Queue. See auto-keyword photos with your categories.
When you do not need machine learning
Skip a trained model if you add a few products a month and a person can categorize each in seconds, if your categories are already set by the supplier feed, or if the photos are mostly text (labels, documents), which the SigLIP 2 page notes is a different problem. Machine learning pays off when photos arrive in volume, categories are your own, and a trusted confidence lets the sure ones go straight through.
Plans
Picture models need a paid base plan plus the Vision add-on: an additional $49 on Essentials or $99 on Business per workspace per month, and each photo read counts toward the plan's monthly pictures allowance. The Dedicated AI Image Model needs Business with Vision. See the pricing page, the product page, image classification and ecommerce product categorization. You can start free with a text model and add Vision when you are ready.
Common questions
How accurate is machine learning for product image classification?
On 1,989 held-out catalog photos in 40 product types, the AI Image Classifier Model was right on 91.1% with all training data and 83.5% with 5 photos per category. Full figures are on the SigLIP 2 benchmark page.
How many photos do I need to train a product photo classification model?
At least 10 per category, and 20 or more is better. In the benchmark, accuracy rose from 85.9% at 10 photos per category to 89.8% at 50.
Can AI sort product photos into folders automatically?
Yes. From S3-compatible storage, AMT can write sorted copies into <folder>/<category>/, with unsure photos in _needs-review/. Your originals are not moved or changed.
Is AI image tagging for ecommerce the same as keywording photos?
In AMT, yes: it can write your category into each photo's Keywords, Title or Description, read by Lightroom and Bridge. Unsure photos are left untagged and sent for review.
Which image model is best for catalog image categorization?
In our test, SigLIP 2 led on product photos at every training size. DINOv2 base led on 37 pet breeds, a fine-grained task. Test on your own photos before you decide.
Does the model read file names or folder names?
No. It learns from the image itself. Folder names, file names or keywords only tell AMT the category of each training photo.