Scene-Designer learns 3D scene design and critique with critic-guided GRPO
Original title: Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation
Two new datasets train design and critique, and test-time critic guidance further improves generation.
Scene-Designer is an agentic framework for generating complete indoor 3D scenes from natural language by learning both design and critique. Scene-OTA-35K supplies observation-thought-action supervision for design and error correction, while Scene-OTC-15K trains critique of physical validity, functional correctness, and layout coherence. Scene-GRPO uses the critic to improve design decisions and reduce unnecessary trial-and-error. Experiments report better scene quality than prior approaches, and the learned critic beats general-purpose models at critique.
- Scene-OTA-35K provides observation-thought-action supervision for design and correction.
- Scene-OTC-15K covers physical validity, functional correctness, and layout coherence.
- Scene-GRPO uses critic guidance to cut unnecessary trial-and-error corrections.
- Test-time critic guidance improves generation beyond the training signal.
- AI models
- Scene-DesignerScene-Critic
From the original article
1 min read at huggingface.co
Generating complete, coherent, and usable 3D indoor scenes from natural language requires strong design intelligence: the ability to design high-quality scenes and critique their quality. However, existing approaches either rely on general-purpose models lacking domain-specific design experience or learn from static scenes.
By Xuanyu Liu, Hongyan Feng, Miao Pan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.