Amazon showed pipeline with Nova, SAM 3, and Textract for automatic PII redaction in photos

AWS Machine Learning Blog presented a multi-step pipeline for automatically masking personal data in images, controlled by the Amazon Nova model.

How the pipeline works

At the core of the solution is contextual visual reasoning from Amazon Nova, which coordinates the work of several complementary tools instead of relying on a single universal algorithm. Nova directs requests to two specialized systems and combines their results into a single image editing process.

• Pipeline orchestrator — Amazon Nova, contextual visual reasoning
• Segmentation — open Segment Anything Model (SAM 3) from Meta
• SAM 3 deployment — on Amazon SageMaker AI platform, pixel-level segmentation
• Text recognition — Amazon Textract, optical character recognition (OCR)
• Complex cases — fingerprints, identity documents, automobile license plates in arbitrary orientation

Why two additional tools are needed

SAM 3 from Meta handles pixel-level segmentation — precise highlighting of image areas that need to be masked, down to non-standard objects like a fingerprint on a document. Amazon Textract, in turn, recognizes text in images, finding fragments like document numbers or signatures that shape-based segmentation might miss. Nova ties these two signal sources together into a single decision about which exact areas need to be edited, ensuring the result is both complete and legally compliant.

The authors specifically emphasize that the pipeline is designed for rare and complex cases — images where personal data is located outside the standard template: at an angle, partially obscured, or belonging to unusual data types such as fingerprints.

What this means

The combination of Nova's visual reasoning with Meta's open segmentation SAM 3 and Textract's OCR engine demonstrates how modern compliance pipelines are increasingly built not on a single model, but on orchestration of several specialized systems, each closing its own class of errors.