Scene-Designer: Critic-guided GRPO teaches agents 3D scene design intelligence
OmniAI-ZJU · hf · 2026-10-19
Zhejiang University's OmniAI Lab introduced Scene-Designer, an agentic framework generating complete, coherent, usable 3D indoor scenes from natural language. It builds two datasets — Scene-OTA-35K (Observation-Thought-Action supervision for design and error correction) and Scene-OTC-15K (Observation-Thought-Critic supervision for critiquing physical validity, functional correctness, layout coherence) — and proposes Scene-GRPO, a Critic-guided GRPO framework that improves design decisions and cuts trial-and-error. It outperforms prior methods on all three metrics; the learned Scene-Critic beats general-purpose models and boosts generation at test time, with design and critique abilities mutually reinforcing.
More from Multimodal
- Training Krea2 LoRA on 16GB VRAM without block swapping, down to 5-7s/it — Corgice · 2026-10-19
- MiniMax Launches Open Multimodal Video Model H3 and M3.1-Flash With 1M Context — testingcatalog · 2026-10-19
- Local Qwen3 27B generates a full video in ~2 hours, completely free — aziz4ai · 2026-10-19
- selflift-Fish: 20s 1080p MiniMax H3 video on a 16GB GPU, 2.2x faster — Hot_Store_5699 · 2026-10-19
- AI-generated series CTVA releases Chapter 3: Disco Fever — cranialtrauma · 2026-10-19
- CTVA AI series drops Chapter 4: The Tallahassee Carrousel — cranialtrauma · 2026-10-19