RoomDreaming: Generative-AI Approach to Facilitating Iterative, Preliminary Interior Design Exploration

Document Title

RoomDreaming: Generative-AI Approach to Facilitating Iterative, Preliminary Interior Design Exploration

Document Information

• Subject Area: Human-Computer Interaction, Generative Artificial Intelligence, Interior Design Assistance Tools
• Keywords: Generative AI, Interior Design, Human-Computer Interaction, User-Centered Design, Design Iteration

Research Background and Problem

• Identified Problems or Challenges:
  • Preliminary interior design exploration often requires multiple meetings and revisions between designers and clients, taking weeks or even months.
  • Homeowners struggle to find reference designs that match their actual spatial layouts using existing tools (e.g., Pinterest, Google Images) and must imagine the combined effects of multiple ideas.
  • Designers often fail to fully understand client preferences, as verbal communication alone frequently leads to misunderstandings.
• Significance:
  • Improving efficiency and optimizing communication between homeowners and designers is a critical need in the field of interior design.
  • Providing more design options in the early stages can help align designs with client needs, reducing rework in later stages.
• Motivation and Related Work:
  • Some generative AI tools for interior design (e.g., RoomGPT, InteriorAI) are available on the market, but they lack the ability to support user preference selection or design iteration.
  • CAD tools can assist in automating design tasks but provide insufficient support for early design exploration.
  • This study aims to combine AI's rapid generation capabilities with user-centered interaction design methods to improve the initial design exploration phase.

Solution

• Method or Solution:
  • RoomDreaming is a generative AI-based tool designed to generate design proposals that match actual room layouts based on user preferences, supporting rapid and extensive iterations.
  • The system includes a web interface, a backend generation module (image analysis, prompt generator, design generator), and a large language model (GPT-3.5).
  • Users can express desired and preferred design directions through "Likes," "Bookmarks," and typed requirements.
  • The system allows users to adjust the balance between new design directions and existing preferences, dynamically controlling the divergence and convergence of designs.
• Innovative Aspects of the Solution:
  • Human-AI Collaborative Generation: Enables homeowners and designers to explore hundreds of design directions more efficiently through iterative generation and filtering.
  • Spatial Layout Matching: AI-generated images closely align with actual room layouts, enhancing the user visualization experience.
  • Adjustable Design Control: Allows users to find a balance between divergent exploration and convergent refinement of designs.
• Key Technologies and Implementation Steps:
  • Image Analysis Module: Uses segmentation models (UPerNet) and depth estimation models (Monocular Depth Estimation) to analyze input room photos, extracting room elements and spatial information.
  • Design Generation Module: Employs ControlNet and Stable Diffusion to generate high-quality images that conform to room layouts.
  • User Preference Integration: Leverages a large language model to analyze user "Likes" and "Bookmarks," generating new prompts to guide design generation.

Research Outcomes

• Specific Results:
  • In five user studies (involving 18 homeowners and 20 designers), RoomDreaming successfully increased the breadth and depth of design exploration while reducing the time required for traditional design iterations.
  • Users were able to complete hundreds of design iterations within one hour; designers estimated this saved 4 to 31 working days compared to traditional design processes.
  • Users expressed general satisfaction with the efficiency and flexibility of the generated designs in the early stages.
• Advantages:
  • Compared to existing generative AI tools, RoomDreaming significantly enhances iteration capabilities and user control.
  • The system's generated designs dynamically reflect room layouts and user preferences, gradually transitioning from divergent exploration to convergent satisfaction.
• Experimental or Evaluation Results:
  • In image quality evaluations, designers indicated that AI-generated images met discussion-level standards in terms of structural systems, functionality, compatibility with user needs, and aesthetic criteria.
  • The system received "Good" to "Very Good" ratings (70%-90%) in both unfurnished and furnished room scenarios.
• Limitations and Future Directions:
  • Some AI-generated designs exhibit limitations in spatial proportions and ergonomics, potentially misleading users regarding practical feasibility.
  • The system currently does not support multi-room design exploration and needs to be expanded to support home-level design (e.g., "HomeDreaming").
  • Plans include developing more creative control features to support bolder or non-traditional design explorations and incorporating regional adaptations to meet the needs of localized users.