Amazon Nova redacts PII from images using vision AI and specialized tools

PII redaction in real-world image datasets presents challenges that traditional masking tools struggle to solve. Unlike structured text, sensitive information in images can appear in unexpected places and forms—a partial face at the edge of a frame, a reflection on a shiny surface, or a partially visible document that becomes identifiable when combined with other visual cues. Amazon's published solution addresses this by positioning Nova 2 Lite as the intelligent coordinator of the entire workflow, using its contextual vision reasoning to assess what constitutes PII in each image before routing analysis to specialized tools.

The architecture separates textual and visual PII detection into parallel processes, allowing Nova to make routing decisions based on its initial assessment. This design reduces cost by preventing unnecessary downstream processing: most business images contain no PII, so an early-exit decision by Nova avoids expensive invocations of Textract and the SAM 3 segmentation model on thousands of images that require no redaction. When PII is detected, the pipeline coordinates pixel-level precision redaction while preserving the image's overall value for downstream use. The solution assumes organizations already have AWS infrastructure in place and are willing to incur charges across multiple services; the body does not state pricing but emphasizes the importance of understanding regional costs before deployment.