Skip to main content

Synthetic Data from a One-Faced Asset

Synthetic reference images let an item be recognised before anyone has photographed it on a shelf. A one-faced asset is the lightest-weight way to get there: you supply a single flat image of the product's packaging artwork, and ZIA generates synthetic reference images from it.

This solves the cold-start problem. A newly listed product has no shelf photography and therefore no reference images, so it cannot be recognised — but its artwork usually exists long before the product reaches a store.

When to use it​

Use a one-faced asset when:

  • You are onboarding a new product and only have pack artwork or a product thumbnail
  • You need recognition coverage from day one of a listing or campaign
  • Photographing the product is impractical, or the product isn't in stores yet
  • You have a large batch of items to onboard and per-item photography isn't realistic

A full 3D asset produces better and more varied synthetic data, because it can be rendered from any angle. A one-faced asset is the pragmatic option when all you have is the front of the pack.

How it works​

  1. Create the catalogue item and its variation.
  2. Submit the artwork image against the variation as a one-faced asset request.
  3. This triggers the synthetic data generation pipeline, which renders the artwork under a range of simulated shelf conditions — different angles, lighting, and surroundings.
  4. The generated images become synthetic reference images for that variation, and are embedded so the models can match against them.
info

One-faced assets are created per variation, not per item. If you are still creating assets at item level, that is the legacy behaviour — new integrations should create the item, then the variation, then the one-faced asset against the variation. Speak to your Neurolabs contact about migrating.

What to supply​

The quality of the synthetic data follows directly from the quality of the artwork:

  • Front of pack, square on. A flat, straight-on view of the primary face.
  • Cropped tightly to the product. No background, no surrounding white space, no other products.
  • High resolution. Detail like flavour names and size markings needs to survive rendering.
  • Correct variant. The artwork must match the variation it is attached to — the whole point is teaching the models to tell near-identical variants apart.

Limitations​

warning

Synthetic data generation is configured per packaging type, and one configuration does not suit everything. A boxed product, a bottle and a can each deform and catch light differently when rendered from a single flat face. For unusual packaging — or for POSM such as posters and strips — check with Neurolabs that the right configuration is in place before onboarding in bulk.

Two further points worth knowing:

Synthetic references are a starting point, not a destination. They get an item recognised on day one. Real crops from actual shelf images consistently outperform them, so items onboarded synthetically should still accumulate real references over time — through Accuracy Improvement or Real Data Flow.

A failed generation looks like a recognition failure. If synthetic data generation doesn't complete or is misconfigured, the item ends up with no reference embeddings at all. The models then have nothing to match against, so genuinely correct detections come back as non-matches. If a newly onboarded item is being missed entirely, check that its synthetic references were actually generated before assuming a model problem.