Parallel Processing Optimization for VITON-HD Virtual Try-On: Performance Analysis and Resource Management Framework

Virtual try-on (VTO) systems face significant computational bottlenecks due to high processing latency and limited scalability, hindering real-time e-commerce applications. This paper presents a CPU-based parallel processing framework that transforms the sequential VITON-HD inference pipeline into concurrent processes using process isolation techniques. Our approach employs a hybrid execution strategy with ProcessPoolExecutor for isolated parallel processes and ThreadPoolExecutor as a fallback mechanism, ensuring system reliability across different deployment environments. The framework incorporates mathematical performance modeling using Amdahl's Law to predict theoretical speedups and guide resource allocation decisions. We implement resource-aware worker scheduling that dynamically adjusts based on available system resources to prevent overload while maximizing throughput. Experimental results on a 12-core system using 2 parallel workers demonstrate a 1.04× speedup with 52.1% parallel efficiency, reducing processing time by 6.7 seconds for batch operations. While the observed performance gains are below theoretical predictions due to process creation overhead and I/O limitations, the framework achieves 100% success rate in batch processing and provides a practical solution for scalable virtual try-on systems without requiring expensive GPU infrastructure. The system offers significant improvements in CPU utilization (78% vs 65%) and enables concurrent processing of multiple person-garment combinations, making VTO technology more accessible for commercial deployment.