FlairGPT: Repurposing LLMs for Interior Designs
EUROGRAPHICS'25
We present layouts generated FlairGPT across a variety of prompts. These range from traditional bedroom and living room designs to more specialized spaces, such as a sewing room, and stylized concepts like "A small workroom for a wizard". FlairGPT also demonstrates its ability to meet specific client-driven functional and aesthetic requirements, such as “A bedroom that is 5x5 for a young girl who likes to paint whilst looking out of her window" or "An office for a bestselling writer in New York who likes to write Fantasy books".
Method overview
Method overview. FlairGPT begins by taking the user's design request as a text prompt and querying an LLM to extract key room parameters, such as dimensions and the location and number of windows, doors, and sockets. Next, following a designer's workflow, the LLM generates an ordered list of zones, specifying the functional purpose of different areas within the room. Based on these zones, a prioritized list of required objects is generated, complete with descriptions and dimensions. These objects serve as the nodes of a layout graph, with inter- and intra-object constraints—defined by the LLM—forming the edges. The natural language constraints provided by the LLM are translated into algebraic forms by querying the LLM to map these constraints to a predefined library of cost functions. Once these cost functions are established, the placement and orientation of objects are progressively optimized according to their hierarchical importance. Finally, objects are retrieved based on their descriptions and incorporated into the scene. We use GPT-4o in all our experiments.
Generated layouts by FlairGPT. We present varied layouts designed by FlairGPT for three distinct prompts (from left to right)- “4m x 5m bedroom", “small workroom for a wizard”, and “bedroom for a vampire”. Alongside each layout, we include descriptions of selected objects provided by the LLM, which closely align with the user’s design brief. Notably, FlairGPT makes creative and context- appropriate object choices, such as a scroll holder and a crystal ball for a wizard’s workroom, and a coffin in place of a traditional bed in the case of a vampire’s bedroom, reflecting the thematic style of the input prompts.
Comparison with other methods
Comparison of layouts generated by FlairGPT against baseline methods, highlighting differences in object arrangement, spatial organization, and overall design quality. We convert room plans generated by LayoutGPT to complete layouts for easier visualization by giving it the same object attributes and room style as ours. We see that layouts generated by LayoutGPT suffer from functional defects whcih lead to blockage of pathways (particularly near the doors). Holodeck and I-Design are limited in stylization and do not offer optimal spatial organization leading to wastage of floor space and non-functional designs.
Select a scene.
Diversity
FlairGPT generates diverse scenes for the same input prompt— “A 5m x 3m home office”, producing a wide range of layouts driven by variations in the selection of objects and style (guided by the LLM), and placement of windows, doors, and sockets. These elements significantly influence the arrangement of objects during our optimization phase, resulting in diverse and dynamic room configurations.
Quantitative Comparison
We compare FlairGPT with closed-universe method— LayoutGPT [FZF*24] and open-universe LLM-based layout generation methods— Holodeck [YSW*23] and I-Design[CHS*24]. The comparison is based on the three metrics outlined in subsection 5.1 of our paper which measure the functionality of the generated layouts in terms of object accessibility (OOB), object overlap (OOR), and access pathway. FlairGPT significantly outperforms all three methods.
Statistics for Experiments
Statistics for our experiments including: the number of primary (P), secondary (S), and tertiary (T) objects per scene; the number of constraints before cleaning, after cleaning, and after translation (function calls); the number of errors including Language errors, Cleaning errors, Translation errors, and Optimization errors; and the time (minutes) for the Language and Translation phase combined, the Optimization phase, and the total time to generate each layout.
Acknowledgements
We thank Rishabh Kabra, Romy Williamson, and Tobias Ritschel for their comments and suggestions. NM was supported by Marie Skłodowska-Curie grant agreement No.~956585, gifts from Adobe, and UCL AI Centre.