In traditional AI-based inspection, deploying a model on a new product requires collecting and labeling hundreds or thousands of defect examples from that specific product. This creates an unavoidable gap between product launch and inspection readiness.

Zero-shot inspection eliminates this gap by transferring defect knowledge learned from one product to a new one, without requiring any labeled defect data from the target product. Here is how it works, what performance to expect, and when it is the right approach.

The new product introduction problem

When a new manufacturing line goes live, the inspection system faces a cold-start problem. The AI model was trained on data from previous products. The new product has different surface characteristics and zero defect examples.

Manufacturers typically handle this in one of three ways: they run manual inspection until enough data accumulates (expensive, error-prone), they delay AI deployment until a sufficient training set exists (leaving a quality gap), or they use rule-based vision systems that cannot handle the variability of real production (brittle, high false-positive rates).

Zero-shot inspection offers a fourth option: deploy AI immediately using knowledge transferred from previous products.

How zero-shot inspection works

The core idea is knowledge transfer. A model trained on defects from an existing product (Product A) learns general defect characteristics. This learned knowledge can be adapted to a new product (Product B) even though no defects have been seen. The adaptation uses synthetic defect generation. Defect representations learned from Product A are placed onto clean images of Product B using generative AI, creating a synthetic training set that teaches the model what defects would look like on the new product. The model is then fine-tuned on this synthetic data before deployment.

Performance expectations

Zero-shot inspection will not match the accuracy of a model trained on thousands of real examples from the target product. But it dramatically outperforms having no inspection at all which is the realistic alternative during new product introduction.

In practical deployments, zero-shot models have achieved approximately 85% mean average precision on new product surfaces. As real defect data accumulates in production, the model can be retrained and improved continuously. The zero-shot deployment serves as the starting foundation.

When to use zero-shot inspection

Zero-shot is most valuable in three scenarios.

New product introduction. A new product is launching and no defect data exists. Zero-shot enables day-one inspection rather than a multi-week or multi-month data collection phase.

High-mix, low-volume manufacturing. Factories producing many product variants in small batches cannot afford to collect dedicated training data for each variant. Zero-shot allows the same defect knowledge to be adapted across the product portfolio.

Rapid line changeovers. Production lines that switch between products frequently benefit from models that can adapt quickly without retraining from scratch each time.

When not to use zero-shot

If you have a stable, high-volume product with ample defect data already collected, a model trained on real data will always outperform a zero-shot model. Zero-shot is a bridge — it is most valuable when real data does not yet exist or is insufficient.

Similarly, for safety-critical inspections where the cost of a missed defect is extremely high (medical devices, aerospace structural components), zero-shot should be used with additional safeguards — such as hybrid human-AI inspection — until the model has been validated against real production data.

How ReliVision enables zero-shot deployment

ReliVision integrates zero-shot capability directly into the platform workflow. Manufacturers can save learned defect concepts from any product, adapt them to new surfaces using on-prem generative AI, and deploy the resulting model to production all without coding or cloud connectivity.