LaviGen is an autoregressive 3D layout generation framework that repurposes the geometric priors of pretrained 3D generative models to produce physically plausible scene layouts directly in native 3D space. Existing text-driven methods typically generate JSON-style layout parameters in a language space and rely on rendering or iterative correction, yet still struggle to ensure physical plausibility and efficiency. In contrast, LaviGen updates the scene state object-by-object with identity-aware encoding and dual-guided self-rollout distillation, improving physical plausibility while reducing computation time and supporting layout completion and editing.
Stay tuned!
- Release technical report
- Release training and evaluation dataset
- Release code and model