A REPRODUCIBLE HARDWARE-AWARE BENCHMARK OF STABLE DIFFUSION V1.5 FOR TEXT-TO-IMAGE GENERATION ON CPU-CLASS SYSTEMS

Abstract

Stable Diffusion generates high-resolution images from text via an iterative denoising process, but incurs large inference time costs on resource-poor computational platforms. This paper shows a controlled hardware-aware benchmark of Stable Diffusion v1. Let us take a deeper look into the stability of text-to-image generation outputs on two CPU-class systems: AMD Ryzen 5 4600H (6C) and AMD Ryzen 7 5800H (8C) Num. It uses classifier-free guidance with a scale of 8, the DDPM sampler, generates 512×512 pairs and measures with seed control ×2. CLIP scores of at least 28% in an eight prompt semantic-alignment test, with a mean score (between +3 to -4) of 31.25%. For latency analysis, three prompts were selected for evaluation with three seeds (42, 200 and 999), which returned nine pairwise observations per hardware comparison. Mean rendering time decreased from 5.35 min on the Ryzen 4600H system to 3.87 min on the Ryzen 5800H system, an absolute saving of 1.48 min per generation and a 27.6% reduction. A paired analysis at the prompt-seed observation level also shows a large effect (Cohen’s dz = 2.81; paired t-test p < 0.001; Wilcoxon signed-rank p = 0.0039). The results indicate that CPU configuration materially affects the practical deployment cost of text-to-image diffusion inference under otherwise consistent settings. The study contributes a reproducible small-scale benchmark protocol and identifies the measurement factors that should be reported when evaluating diffusion-model latency on heterogeneous systems.