Emotional preferences in LoRA-driven AI-generated urban vitality spaces: A study using discrete choice models

China's transition toward urban stock renewal has heightened the focus on the emotional quality of urban vitality spaces. While AI generative design offers new approaches, whether AI-generated images as visual representations for specific spaces can meet people's emotional needs remains insufficiently examined. This study examines whether AI-generated visual representations of such spaces evoke image-based emotional preferences comparable to those for real photographs, and identifies which spatial attributes drive these preferences. Using THE BOX in Beijing as a case, Low-Rank Adaptation (LoRA)-fine-tuned Stable Diffusion models generated visual representations of seven chosen scenes. Emotional preferences were measured through validated scales completed by 115 participants. A discrete choice model was used to compare participants' preferences between photographs of the real space and AI-generated images and identified influential attributes. Results showed that participants with design-related backgrounds preferred real photographs. Positive affect toward AI-generated images was the strongest predictor of preference. Real-scene color, natural elements, and uniqueness reduced the odds of choosing AI-generated versions, whereas AI-generated form and structure, natural elements, and changeability increased them. The study contributes a replicable workflow linking generative AI with emotional preference measurement, offering a structured approach for evidence-based design review. The findings are preliminary due to the single case, modest sample, and reliance on static images. Future work should extend to diverse urban contexts and multimodal evaluation.