Research-Multi-scale controllable scene generation

DiffuTown explores a multi-scale generative pipeline for transforming diffusion-based images into structured and controllable urban scenes. Starting from text prompts, the system generates global layout images and progressively refines them into localized regions through semantic segmentation, focus-based cropping, and controlled diffusion. Each stage introduces increasing levels of spatial detail, from city-scale planning to architectural and environmental elements. The pipeline integrates vision-language models, segmentation models, and diffusion control mechanisms to extract semantic patches, reconstruct spatial hierarchies, and enable consistent layout generation across multiple scales. By bridging image generation with spatial structure, DiffuTown reframes diffusion models as a tool for computational urban design — enabling iterative refinement, asset-level control, and potential extension toward 3D scene reconstruction.