Step-by-step guide to compiling and deploying Stable Diffusion XL on AWS Inferentia2 using Amazon SageMaker and the optimum-neuron library. Covers model conversion to Neuron format, creating a custom inference.py script, uploading artifacts to S3, deploying a real-time SageMaker endpoint, and running image generation. The setup achieves ~6 seconds per image on an inf2.xlarge instance at $0.99/hour, translating to roughly $0.0016 per generated image — positioning Inferentia2 as a cost-effective GPU alternative for inference workloads.
