Autoregressive 3D Diffusion model generates scene layout and object shape from text

Researchers have developed a new 3D autoregressive diffusion model called 3D-ARD+ that can generate both scene layouts and object shapes from text descriptions. This model sequentially generates objects, first creating coarse 3D latents in the scene space and then refining them into fine-grained object geometry and appearance. Trained on a dataset of 230,000 indoor scenes, 3D-ARD+ aims to create more consistent and detailed 3D scenes that accurately follow complex text instructions. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a new method for generating detailed 3D scenes from text, potentially improving tools for virtual environment creation and content generation.

RANK_REASON This is a research paper describing a novel generative model for 3D scene creation.

Read on arXiv cs.CV →

• 3D-ARD+
• arXiv
• Zhenggang Tang
• Hugging Face

• paper
• model release