AWS demonstrates object detection using Amazon Nova 2 Lite multimodal model with no training required

AWS published a technical guide showing how Amazon Nova 2 Lite performs object detection through natural language prompts without requiring model training, data pipelines, or dedicated infrastructure.

Pricing and capabilities

Amazon Nova 2 Lite costs $0.0003 per thousand input tokens and $0.0025 per thousand output tokens when accessed through Amazon Bedrock. According to AWS, a typical image consumes approximately 230 input tokens ($0.000069) and generates around 200 output tokens ($0.0005), totaling roughly $0.00057 per image. Processing 10,000 images would cost approximately $5.69.

The model accepts natural language prompts specifying objects to detect (such as "vehicle", "person", or "dent") and returns bounding box coordinates in structured JSON format. Coordinates use a normalized 0-1000 scale that developers convert to pixel positions.

Technical implementation

The implementation uses Amazon Bedrock's Converse API with a prompt engineering template that specifies detection requirements. The prompt includes two dynamic variables: elements (object types to detect) and schema (expected JSON structure). The system requires no fine-tuning or training data.

AWS tested the model on a street scene, asking it to detect "vehicle" and "stop sign" objects. According to AWS, Nova 2 Lite detected small, distant, and partially occluded objects with tight bounding boxes using only basic object names.

Architecture and deployment

AWS released a reference serverless application architecture combining AWS Lambda, Amazon API Gateway, Amazon CloudFront, and Amazon S3. The Lambda function orchestrates requests to Amazon Bedrock, converts normalized coordinates to pixel positions, and renders bounding boxes on images.

The architecture supports deployment on AWS Lambda for event-driven workloads, Amazon EC2 for custom configurations, or Amazon ECS/EKS for containerized deployments. All compute options use the same Bedrock Converse API.

AWS estimates deployment takes 30-45 minutes and provides complete source code including AWS CDK infrastructure definitions in a GitHub repository.

What this means

Amazon Nova 2 Lite offers a low-cost alternative to traditional computer vision pipelines that require data collection, model training infrastructure, and ML expertise. At under $0.001 per image, the model makes object detection economically viable for small teams and prototyping scenarios. The prompt-based approach eliminates training costs but likely sacrifices accuracy compared to domain-specific models trained on custom datasets. The reference architecture demonstrates production-ready deployment patterns, though AWS has not published benchmark comparisons against established computer vision models or disclosed Nova 2 Lite's base parameter count.