Amazon shows Nova-driven pipeline using SAM 3 and Textract to automatically redact PII in images

The AWS Machine Learning Blog has presented a multi-step pipeline for automatically redacting personal data in images, with the Amazon Nova model in control.

How the pipeline works

At its core is Amazon Nova's contextual visual reasoning, which coordinates several complementary tools instead of relying on a single universal algorithm. Nova routes requests to two specialized systems and merges their results into one image-redaction workflow.

• Pipeline orchestrator — Amazon Nova, contextual visual reasoning
• Segmentation — Meta's open Segment Anything Model (SAM 3)
• Where SAM 3 runs — on Amazon SageMaker AI, pixel-level segmentation
• Text recognition — Amazon Textract, optical character recognition (OCR)
• Difficult cases — fingerprints, ID documents, license plates in arbitrary orientation

Why two additional tools are needed

Meta's SAM 3 handles pixel-level segmentation: it precisely outlines the areas of the image that must be hidden, down to unusual objects such as a fingerprint on a document. Amazon Textract, in turn, recognizes text in the image and finds fragments such as document numbers or signatures that object-shape segmentation could miss. Nova combines these two signal sources into a single decision on which areas to redact, so that the result is both complete and legally compliant.

The authors specifically stress that the pipeline is built for rare and difficult cases: images where personal data does not follow a standard layout, such as data placed at an angle, partly obscured, or belonging to atypical data types like fingerprints.

What this means

Combining Nova's visual reasoning with open-source SAM 3 segmentation and the Textract OCR engine shows how modern compliance pipelines are increasingly built not on a single model but on the orchestration of several specialized systems, each covering its own class of errors.