One-pass 3D scene synthesis models often produce spatial hallucinations such as object collisions because they lack deliberative reasoning. SceneReVis, a framework introduced by researchers from Shanghai Jiao Tong University and Microsoft Research, addresses this by embedding a vision-grounded self-reflection loop that explicitly diagnoses and resolves spatial conflicts before finalizing a scene.

To support this step-wise paradigm, the team constructed SceneChain-12k, a large-scale dataset of causal construction trajectories derived through a reverse engineering pipeline. They then trained the model in two stages: first with supervised fine-tuning, then with agentic reinforcement learning, which evolves the system into an active spatial planner rather than a one-shot generator.

In experiments, SceneReVis achieved state-of-the-art performance in high-fidelity generation and goal-oriented optimization. It also transferred better to underrepresented room categories and showed promise for out-of-distribution room types. The paper has been accepted as an oral presentation at ACM Multimedia 2026; this digest is based solely on that source, with no conflicting reports to compare.