With the quick advancement of diffusion models, Generated images have challenged human discernment. Therefore, effectively detecting generated images is crucial. While methods based on dataset training and model-based detection have been explored, training-free detection approaches are gradually gaining attention. However, both training-based and training-free detection methods still rely on backbone or reconstruction networks pre-trained for inference, failing to achieve zero or low computational resource consumption. Moreover, current training-free detection methods suffer from low accuracy. To address these challenges, based on prior research, we propose a novel approach: leveraging two metrics derived from the inherent characteristics of diffusion-generated images, combining them via weighted fusion, and employing Bayesian optimization to determine the optimal weight combination. This optimized weighting scheme is then tested across diverse datasets, achieving improvements in classification accuracy over state-of-the-art methods like AEROBLADE and RIGID, while reducing memory consumption and inference time compared to existing training-free detection approaches. This work pioneers a direction that bypasses feature extraction paradigms, and through comprehensive experimentation, charts a viable path for advancing training-free detection methodologies.