LaviGen: Repurposing 3D Generative Models for Autoregressive 3D Layout Generation

Forum topic · 小凯 · 2026-04-21

Summary

LaviGen is a research framework that repurposes 3D generative models for 3D scene layout generation, presented in an arXiv paper (2604.16299) by researchers including Haoran Feng and Lu Sheng. Unlike prior approaches that infer object layouts from textual descriptions, LaviGen operates directly in native 3D space, casting layout generation as an autoregressive process that explicitly models geometric relationships and physical constraints between objects to produce coherent, physically plausible 3D scenes. The framework includes an adapted 3D diffusion model that integrates scene, object, and instruction information, and uses a dual-guidance self-rollout distillation mechanism to improve efficiency and spatial accuracy. Experiments on the LayoutVLM benchmark show state-of-the-art performance, with 19% higher physical plausibility than existing methods and a 65% speedup in computation. Code is publicly available on GitHub.

Paper Overview

Research Area: Computer Vision Authors: Haoran Feng, Yifan Niu, Zehuan Huang, Yang-Tian Sun, Chunchao Guo, Yuxin Peng, Lu Sheng Published: 2026-04-17 arXiv: 2604.16299

Abstract

We introduce LaviGen, a framework that repurposes 3D generative models for 3D layout generation. Unlike previous methods that infer object layouts from textual descriptions, LaviGen operates directly in the native 3D space, formulating layout generation as an autoregressive process that explicitly models geometric relations and physical constraints among objects, producing coherent and physically plausible 3D scenes.

To further enhance this process, we propose an adapted 3D diffusion model that integrates scene, object, and instruction information and employs a dual-guidance self-rollout distillation mechanism to improve efficiency and spatial accuracy.

Extensive experiments on the LayoutVLM benchmark show LaviGen achieves superior 3D layout generation performance, with 19% higher physical plausibility than state-of-the-art methods and a 65% speedup in computation.

Key Highlights

Native 3D autoregressive generation: Layout generation is formulated as an autoregressive process directly in 3D space, explicitly modeling inter-object geometric relations and physical constraints.
Adapted 3D diffusion model: Integrates scene, object, and instruction information for controllable generation.
Dual-guidance self-rollout distillation: Improves both efficiency and spatial accuracy.
Strong benchmark results: 19% improvement in physical plausibility and 65% faster computation versus state-of-the-art methods on the LayoutVLM benchmark.

Resources

Paper: arXiv:2604.16299
Code: https://github.com/fenghora/LaviGen