I-Design: Personalized LLM Interior Designer

TL;DR

The paper tackles the barrier to personalized interior design by introducing I-Design, a multi‑agent LLM framework that converts unstructured natural language into a structured 3D interior scene. It couples a scene‑graph representation with a backtracking 2D layout generator and a 3D asset retrieval pipeline to produce feasible, visually coherent designs, which are rendered and evaluated via a vision‑language model. The authors validate their approach against LayoutGPT, showing improvements in object plausibility, spatial coherence, and alignment with user intent through quantitative metrics and human‑inspired assessments. The work contributes an interpretable, end‑to‑end design pipeline, a novel VLM‑based evaluation protocol, and extensive experimentation that demonstrates the potential of LLM‑driven, multi‑agent interior synthesis for non‑professional users. It lays groundwork for further improvements in automated, adaptable, and texture‑aware 3D interior design, while acknowledging limitations in termination, asset quality, and texture consistency that guide future work.

Abstract

Interior design allows us to be who we are and live how we want - each design is as unique as our distinct personality. However, it is not trivial for non-professionals to express and materialize this since it requires aligning functional and visual expectations with the constraints of physical space; this renders interior design a luxury. To make it more accessible, we present I-Design, a personalized interior designer that allows users to generate and visualize their design goals through natural language communication. I-Design starts with a team of large language model agents that engage in dialogues and logical reasoning with one another, transforming textual user input into feasible scene graph designs with relative object relationships. Subsequently, an effective placement algorithm determines optimal locations for each object within the scene. The final design is then constructed in 3D by retrieving and integrating assets from an existing object database. Additionally, we propose a new evaluation protocol that utilizes a vision-language model and complements the design pipeline. Extensive quantitative and qualitative experiments show that I-Design outperforms existing methods in delivering high-quality 3D design solutions and aligning with abstract concepts that match user input, showcasing its advantages across detailed 3D arrangement and conceptual fidelity.

Figures (8)

• Figure 1: Overview of I-Design. Starting from the user specification of design preferences in plain text, we query LLM agents to come up with room items, their properties, and their relative relationships in the form of a scene graph. We solve for absolute object placement in the scene graph using the proposed backtracking algorithm (Scene Graph Layout), retrieve 3D assets according to the functional and stylistic specifications, and compose the final result in 3D.
• Figure 2: Scene Graph Generation Pipeline with LLM Agents. Each agent receives the user input and a JSON representation of the previous stages and transforms the output according to the task prompt. Due to the specialized nature of each agent, the generated scene graph reflects user specifications and feasibility constraints, such as topological correctness and semantic plausibility. Red indicates changes introduced by each agent. Complete prompts of every agent are specified in the Appendix.
• Figure 3: Backtracking-based Scene Graph Layout. Starting from the scene graph specifying interior items as nodes connected with relative relationships, the backtracking algorithm solves for their absolute placement in the room. The algorithm reverses dead-end configurations and excels at object placement while respecting spatial constraints.
• Figure 4: Gallery of Results Obtained with I-Design. The first column lists user prompts, specifying design preferences, such as functionality, style, and atmosphere. The top-down scene graph layout generated by the LLM Agents is shown in the second column. The third column shows this layout rendered from the top-down view after the Asset Retrieval stage. The last two columns show corner views of the resulting design. Evidently, I-Design is capable of factoring in diverse user specifications and producing vibrant, functional designs.
• Figure 5: Control Questions for the Subjective Study Here are two control questions provided for our subjective study focusing on bedrooms in Figure \ref{['fig:userstudybed']} and living rooms in Figure \ref{['fig:userstudyliv']}. These questions are designed to prompt participants to assess object collisions and the level of detail present in the scenes. The red boxes show the object collisions in the scenes.