DiffDesign: Controllable diffusion with meta prior for efficient interior design generation

Area of Science:

• Artificial Intelligence
• Computer Vision
• Computational Design

Background:

• Interior design is complex, requiring balancing aesthetics, functionality, and ergonomics.
• Current generative models lack control and practical relevance for interior design tasks.
• Existing methods struggle with size, spatial scope, and generation quality.

Purpose of the Study:

• To develop an efficient and controllable generative model for interior design.
• To address the limitations of existing generative approaches in interior design.
• To improve the practical applicability of AI in interior design generation.

Main Methods:

• Proposed DiffDesign, a controllable diffusion model leveraging meta-priors.
• Utilized a pre-trained 2D diffusion model as a rendering backbone.
• Implemented cross-attention control for design attributes (appearance, pose, size) and an alignment module for view consistency.
• Constructed the DesignHelper dataset (400+ solutions, 15+ spatial types, 15+ styles) for fine-tuning.

Main Results:

• DiffDesign demonstrates effective and robust interior design generation.
• The model shows improved controllability over design attributes.
• Ensured view consistency in generated designs.
• Validated through extensive experiments on benchmark datasets.

Conclusions:

• DiffDesign offers an efficient solution for AI-powered interior design.
• The controllable generation capabilities enhance practical usability.
• The developed dataset and model advance the field of computational interior design.