Cot2Scene: Generating 3D Indoor Scenes Using CoT-Enhanced LLMs

Current indoor scene generation algorithms face significant limitations in alignment with user instructions and ensuring logical scene coherence. To address these challenges, we propose CoT2Scene, a framework that leverages Chain-ofThought (CoT) to enhance LLMs for 3D indoor scene generation. First, user instructions are processed by the LLM and the closest matched layout examples are retrieved from a knowledge base. Then, with the retrieved examples, the framework queries the LLM using a meticulously designed CoT template to generate spatial layouts that are both intent-aligned and logically structured in a step-by-step manner. Subsequently, high-quality 3D models are retrieved from the 3D-FUTURE dataset to construct the final 3D indoor scene. Experimental results demonstrate that, compared to prior methods, our approach not only generates scenes that closely align with user intent but also outperforms existing methods in terms of scene quality and logical coherence.