Generating a 3D indoor scene from a single image is useful for robotics, embodied AI, and virtual environments, but current methods tend to produce layouts that look plausible yet break physical rules—objects floating, intersecting, or placed where they could not function. The authors of this preprint propose a Physics Evaluator that scores scenes on four aspects: geometric priors, contact, stability, and deployability, broken down into nine sub-constraints. Using this evaluator as a benchmark, they find that even state-of-the-art generators produce scenes with substantial physical inconsistencies.

To address this, the paper introduces PhyMix, a framework that feeds evaluator feedback into both training and inference. It has two components: implicit alignment via Scene-GRPO, a critic-free group-relative preference optimization that nudges the generator toward physically plausible layouts, and explicit refinement via verifier-guided agentic repair, which translates evaluator diagnostics into concrete scene-editing actions at inference time.

The authors report that PhyMix improves both reconstruction quality and physical plausibility compared with existing methods. Qualitative examples on stylized and real-world images suggest the approach generalizes beyond synthetic inputs. As a single preprint, the results have not yet been independently replicated, but the unified evaluation-plus-correction strategy offers a clear direction for making generated scenes more physically reliable.