Episode 4 · 7:42
Batch AI fashion photography with automatic QA
A batch applies the same settings to many products at once, and automatic QA checks every photo it makes. This episode sets up one batch for several outfits, with a front photo and two camera edits each, switches on the quality check, and shows what happens when a photo fails: the agent flags a jacket that came out the wrong color and explains why.
Chapters
What you'll learn
- How a batch is built: a first photo, then edits on it
- How to shoot single garments or outfits you already built
- What automatic QA checks, and what a failed check looks like
- How approve, reject and retries work for a team
How a batch works
A batch has two steps. Step 1 makes the first photo of each product: the model, the image model, the art direction and the camera angle. Step 2 adds operations to that photo, such as camera edits for the side and the back.
You see the full list of photos and their credits before anything runs, so a batch of fifty products has a known price before you start it. To shoot every new product as it arrives instead, use an automation.
Choose the inputs
A batch can take single garments or outfits you have already built, for example a jacket with the trousers it is sold with. Pick them from your library.
Set the first photo and the edits
Set the first photo once for the whole batch: the model, the image model, the art direction and the angle. Then add two camera edits to every first photo, the side and the back.
To shoot the same products a second way, change one setting, such as the model, and add a second group to the same batch. That is how one run can cover several models for several markets.
Switch on automatic QA
At volume, nobody looks at every photo. Switch on QA and an agent compares each photo with its inputs: the garment and the model. It checks what a person would check, the color, the details, the fit and whether it is the right model, and it writes down why a photo fails.
The check is free. You can also add retries, up to five: a photo that fails is shot again with the same request, and each retry costs the same as a new photo.
What a failed check looks like
To show a failure up close, the episode makes one on purpose: the prompt turns a dark brown jacket blue. The agent flags the color and explains the mismatch with the garment photo. A passed check is shown too, with the list of checks it passed.
A photo that still fails after its retries is marked rejected. Nothing is thrown away: it stays on the record, and a person can approve it anyway.
Approve and reject as a team
Approve and Reject let your team mark results by hand, on top of the agent's check. The Generations page filters by QA result, so the photos that need a human look are easy to find, and the rest can go straight to your store.