Aesthetic-driven and sketch-guided dual-conditioned latent diffusion model for urban brand image synthesis

Abstract

Urban brand image design requires regional cultural symbols to be preserved while visually appealing content is generated efficiently. General text-to-image diffusion models are useful for visual content production, but they often provide limited spatial control and may inadequately represent local cultural motifs. This paper proposes an aesthetic-driven and sketch guided dual-conditioned latent diffusion model (AD-SG-LDM) for urban brand image synthesis. The framework uses a latent diffusion backbone for text-guided generation, a ControlNet branch to transform weak sketches into spatial constraints, and low-rank adaptation (LoRA) modules to adapt the frozen model to a small Xiamen cultural-image dataset. An aesthetic guidance term based on palette consistency and composition energy is further incorporated during reverse sampling to bias generation toward visually coherent results. A dataset of 2,000 Xiamen-related images was constructed and annotated for experimental evaluation. Ablation results show that sketch control, LoRA adaptation, and aesthetic guidance progressively improve FID, CLIP Score, and human aesthetic ratings. The full model achieves an FID of 12.38, a CLIP Score of 0.345, and a mean user-study score of 4.78/5, outperforming the tested baselines. The results indicate that combining spatial constraints, domain adaptation, and aesthetic guidance can improve controllability and visual quality in regional cultural image generation.