How to train an image classifier without code, using photos you've already sorted

An image classifier looks at a whole picture and gives it one of your categories, with a confidence. Think of a person glancing at a photo and saying "that's water damage".

Teams already make these calls every day:

The categories that matter are your categories, not a generic list of objects. That's why you train on your own sorted photos instead of using a ready-made "what's in this picture" AI.

• Photos you've already sorted, or can sort quickly: at least 10 per category, ideally 20–50.
• Photos like the ones you'll classify later: the same kind of camera, lighting, distance and background. A model trained on studio shots will struggle with phone photos taken on a loading dock.
• A clear rule for each category that two people would apply the same way. If your team argues about "crushed" vs "dented", the model will be unsure too.
• StayCharted AI Model Trainer with Vision. Vision is an add-on on the Essentials and Business plans.

1. Pick one decision

One question, a handful of answers: "What condition is this parcel in?" Don't mix questions ("condition" and "carrier") in one list of categories. Train two models instead.

2. Collect representative photos

Pull photos from where they already are: your WMS, claims system, shared drive or a phone's camera roll. Include the hard cases: a crushed corner that's only slightly crushed, water damage on a dark box. Those are what the model needs to learn.

3. Make one folder per category

The folder name becomes the category:

parcel-photos/
  Intact/
  Crushed corner/
  Torn tape/
  Water damage/

The model learns from the picture itself, never from the file name. "damaged_001.jpg" gives it nothing to cheat with.

4. Make the photos smaller, then ZIP the folders

StayCharted keeps a 512-pixel copy of each photo, which is what the model reads, so full-size phone photos only make the upload slower.

• Resize to about 512–800 pixels on the long side (Preview or Photos on a Mac, the Photos app or PowerToys Image Resizer on Windows).
• Use JPEG, PNG or WebP. iPhone HEIC photos need converting to JPEG first.
• ZIP the top folder. If the ZIP is over your plan's limit, use smaller photos, or upload a spreadsheet with a link to each photo and its category instead.

5. Create a picture model and upload

In StayCharted, create a model, choose pictures, and upload the ZIP. It reads every photo in the background, and you can leave the page while it does. Location data and other hidden details are removed from each picture on upload.

6. Check your pictures

This is the step that most improves a model, and it takes minutes. StayCharted shows you:

• photos that look like they belong in another category, with a suggestion;
• the same photo in two folders;
• near-identical shots of the same parcel;
• files that couldn't be read.

Each finding is one click to fix: move it, keep it, or leave it out. Click any thumbnail to see it larger.

7. Train, and read the results

Train an AI Image Classifier Model. It usually takes about a minute. StayCharted holds back part of your photos, trains on the rest, and reports how often the model was right on the photos it never saw.

Read two things beyond the overall score:

• Accuracy per category. A model can be 95% right overall and still miss most of a rare category.
• The mistakes themselves. If "crushed corner" photos keep landing in "intact", look at them. Often they're borderline photos that need a clearer rule or more examples.

8. Try it on a new photo

Open Try a picture and choose a photo from your computer, or paste a link. You get the category, the confidence and a thumbnail of the photo. Click the thumbnail to check it larger. Mark the answer right or wrong; that feedback teaches the next version.

This is the step to show your team: a photo nobody has labeled, and the answer next to it.

9. Put it to work

• A batch of photos: upload a ZIP of new photos, or a spreadsheet of photo links, and get back a spreadsheet with a category and a confidence for each.
• From your own system: call the model through the API, with a photo link or the photo itself, up to 20 photos a call.
• Unsure answers to a person: set a review cutoff. Answers below it go to the Review Queue, where someone confirms or corrects them.

10. Retrain with what you learn

Corrected answers become training photos for the next version. Retrain when you see new kinds of mistakes, and whenever something changes: a new camera, a new warehouse, a new packaging supplier.

Start with the AI Image Classifier Model. Try the Dedicated one if the mistakes are on small details. In our food grading benchmark:

• Fruit quality: both models were right 98.7% of the time.
• Cashew defects: the Dedicated model found 65% of defective cashews, against 40%.

Compare both on the same held-back photos before you decide. Training the AI itself doesn't guarantee better results.

• Near-identical shots split across training and testing. Five photos of the same parcel from slightly different angles make the test too easy. Remove near-duplicates (step 6) before you trust the score.
• The background gives the answer away. If every water-damage photo was taken in the claims room and every intact one on the dock, the model may learn the room, not the damage. Mix backgrounds within each category.
• One category is much larger than the others. A model that calls everything "intact" can still score well if 90% of photos are intact. Check accuracy per category.
• Old photos, new reality. Retest after anything changes in how photos are taken.

Often the photo isn't the whole story: a customer's message comes with their damage photo, or an inspector adds a note. A picture model can read a text column alongside each photo and learn from both. A row with a note but no photo is answered from the note.

• It doesn't find where the damage is. It classifies the whole photo; it doesn't draw boxes around defects (that's object detection).
• It doesn't read text in the photo. It's not OCR. If the words on a label decide the category, use the text instead.
• It doesn't identify an exact item or SKU. It predicts your categories.
• It doesn't hide faces or visible text. Hidden metadata is removed on upload; what's visible in the photo stays.

Not ready to gather your own photos? In StayCharted, Start here has a parcel sample: 44 photos already in folders for intact, crushed corner, torn tape and water damage. It trains in about a minute and shows every step above.