Generating usable 3D indoor scenes from natural language requires more than object placement; it demands design intelligence—the ability to create coherent layouts and judge their quality. Existing systems typically rely on general-purpose models that lack domain-specific design experience, or they learn from static scenes without feedback. The new Scene-Designer framework instead trains an agent to both design and critique scenes, using two complementary forms of experience.

The authors construct two datasets: Scene-OTA-35K provides observation–thought–action supervision for learning scene design and efficient error correction, while Scene-OTC-15K supplies observation–thought–critic supervision for evaluating physical validity, functional correctness, and layout coherence. On top of these, they propose Scene-GRPO, a critic-guided reinforcement learning method that improves design decisions and reduces unnecessary trial-and-error corrections.

Experiments show Scene-Designer outperforms prior approaches across all three quality metrics. The learned Scene-Critic also beats general-purpose models at critique, and applying the critic during test time further boosts generation quality. The authors observe a trend of mutual enhancement: better design improves critique, and better critique improves design, suggesting that the two abilities reinforce each other during training.