Customers upload images (totaling up to 50 MB each) and an ML model generates textual descriptions. The images are stored in Amazon S3 and the processing must scale with minimal operational overhead. Which architecture best satisfies these requirements?

An Amazon SageMaker Asynchronous Inference endpoint is ideal for this scenario because it handles large payloads (up to 1 GB) and long processing times, which is suitable for image processing and text generation. It automatically queues requests and scales the underlying instances based on the workload, minimizing operational overhead. Invoking it for each image ensures individual processing. SageMaker batch transform is less suitable because it processes an entire dataset in a single job, which isn't ideal for real-time or near real-time processing of individual uploads. Amazon EKS with Karpenter offers high flexibility but introduces significant operational overhead compared to a managed SageMaker endpoint. AWS Batch with ECS is also a viable option for batch processing but typically involves more setup and management than SageMaker Asynchronous Inference for this specific use case of individual, event-driven image processing.