Conditional Diffusion Model Optimization for Real-Time Point-of-Care Image-Quality Enhancement on Nvidia Jetson Orin Nano

High-resolution optical coherence tomography (OCT) is often inaccessible in low-resource settings. While diffusion models can enhance low-cost OCT scans, their computational cost prohibits embedded deployment. We solve this problem by presenting the first real-time deployment of a diffusion-based OCT enhancement model on an Nvidia Jetson Orin Nano. Our systematic pipeline achieves a 33 ms inference time by first pruning the model from 224 M (million) to 42.9 M parameters (80.8% reduction) with minimal fidelity loss (no statistically significant difference in quality (p>0.3) ). We then apply a novel hybrid INT8 quantization approach, strategically offloading unstable Softmax operations to a lightweight custom CUDA kernel. Expert evaluation shows that downstream diagnostic utility of the resulting images is also not impacted. This work provides a concrete blueprint for translating large generative models into practical, affordable medical diagnostic tools.