AWS launches Nova-powered PII redaction pipeline for images using SAM 3 and Textract

AWS deploys Nova-powered PII redaction for enterprise image processing

AWS has introduced an automated pipeline for redacting personally identifiable information in images, using Amazon Nova 2 Lite to coordinate multiple specialized tools. The solution addresses compliance requirements under GDPR and PCI DSS by detecting and obscuring PII in challenging scenarios where single-purpose tools typically fail.

The pipeline uses Nova 2 Lite as the central decision-making system, coordinating Meta's open-source Segment Anything Model 3 (SAM 3) deployed on Amazon SageMaker for pixel-level segmentation, and Amazon Textract for optical character recognition.

Architecture and workflow

According to AWS, the system operates through a multi-step process:

1. When an image is uploaded to S3, EventBridge triggers an AWS Step Functions workflow
2. Nova 2 Lite performs initial PII assessment, routing images without PII directly to output to reduce costs
3. For images containing PII, Nova classifies detected information as textual (names, addresses, ID numbers) or visual (faces, fingerprints)
4. Nova selectively invokes specialized processes: Textract for text extraction with coordinates, SAM 3 for visual element segmentation
5. Lambda functions obscure content at identified coordinates

The solution handles edge cases including partial faces at frame edges, reflections on surfaces, partially visible street signs, and documents in wide-angle photos that reveal sensitive information.

Technical capabilities

Nova 2 Lite's multimodal vision understanding enables it to interpret image content holistically and reason about whether elements constitute PII in context. AWS states this contextual reasoning distinguishes the system from single-purpose masking tools.

The pipeline detects multiple PII categories: names, identification numbers, addresses, telephone numbers, MAC addresses, vehicle identification numbers, facial images, and fingerprints.

Pricing for the solution includes S3 storage, Lambda invocations, Step Functions state transitions, SageMaker endpoint hosting, Amazon Bedrock API calls, and Textract API calls. AWS has not disclosed bundled pricing for the complete pipeline.

Deployment requirements

The solution requires:

• AWS account with access to Nova 2 Lite via Amazon Bedrock
• SAM 3 deployment on SageMaker
• Familiarity with Amazon Bedrock, SageMaker, S3, Lambda, Step Functions, EventBridge, and Textract
• Basic computer vision and prompt engineering knowledge

What this means

This release demonstrates AWS's strategy of positioning Nova models as orchestration layers rather than standalone inference engines. By coordinating specialized tools—SAM 3 for segmentation, Textract for OCR—Nova acts as an intelligent router that reduces costs through early-exit decisions while maintaining accuracy for complex cases. The approach suggests AWS views multimodal foundation models as coordinators for multi-tool pipelines rather than end-to-end solutions, a potentially significant architectural pattern for enterprise AI deployments facing strict compliance requirements.