Local Precise Redrawing of Architectural Scheme Renderings Based on Improved Diffusion Models

This paper presents a novel approach to architectural visualization through the development of improved diffusion models capable of local precise redrawing in architectural scheme renderings. We address the critical limitation of current diffusion models in performing targeted, localized modifications while maintaining overall design coherence and architectural accuracy. Our methodology combines advanced diffusion architectures with sophisticated local control mechanisms, including enhanced inpainting techniques, multi-scale attention mechanisms, and architectural domain-specific fine-tuning. Through extensive experimentation on a curated dataset of 10,000+ architectural images, we demonstrate significant improvements in local precision control, achieving a local SSIM score above 0.85 and FID score below 50. Our integrated framework incorporates ControlNet for multi-modal control, LoRA fine-tuning for architectural domain adaptation, and novel loss functions designed specifically for architectural constraints. Human evaluation studies with 15 expert architects and 50 general users validate the practical applicability of our approach, showing professional assessment scores above 4.0/5.0. The proposed system enables architects to perform precise local modifications in seconds rather than hours, fundamentally transforming the iterative design process while maintaining high visual quality and architectural integrity.