Urban streetscape evaluation has traditionally relied on computer vision methods to quantify semantic elements such as greenery, signage, and architectural features. However, these objective approaches often overlook subtle factors like lighting and atmospheric conditions that deeply affect human perceptions of beauty, safety, and ambiance. Moreover, existing expert-driven evaluations may not fully represent diverse public perspectives. To address these limitations, this study proposes a perception-oriented framework integrating generative AI and large-scale subjective feedback to investigate how weather conditions shape urban streetscape perception. Using TSIT-based image translation, we generated sunny, cloudy, and rainy versions of 88,000 street view images while preserving scene structure and semantics. Over 38,525 participants provided subjective ratings across 22 perceptual dimensions, such as beauty, cleanliness, safety, and warmth. Results revealed that sunny scenes scored higher than cloudy and higher than rainy scenes on the most of 22 dimensions, with notable improvements in brightness, neatness, and cleanliness. Rainy conditions elicited the highest negative ratings in dimensions like boredom and depression, while dimensions such as greenery and interestingness remained relatively stable across weather types.