Scene-Designer Learns to Design and Critique 3D Scenes from Text Using 50,000 Examples

Design Intelligence for Text-to-3D Scenes

Turning plain text into complete, coherent, genuinely usable 3D scenes demands a specific skill — what researchers call "design intelligence." That's the ability not just to design scenes, but to judge their quality. Existing approaches either lean on general-purpose models with no specialized design expertise, or they're trained only on static scenarios.

A team of researchers introduced Scene-Designer, an agentic framework that develops both of these complementary abilities simultaneously, drawing on examples from both design and critique. The key idea: the model doesn't just generate 3D scenes — it develops the ability to evaluate them in parallel, creating a closed-loop improvement system.

Two Datasets: Design and Critique

To train the model, the researchers built two specialized datasets:

Scene-OTA-35K contains 35,000 examples aimed at training efficient scene construction and error correction during scene building.

Scene-OTC-15K includes 15,000 examples for training critique of scene quality across three criteria:

• Physical validity — whether objects are placed logically in space and positioned sensibly relative to one another
• Functional correctness — whether the scene matches its intended purpose and objects interact correctly
• Layout consistency — harmonious, balanced placement of all elements

Scene-GRPO: The Critic as a Guiding Force

To optimize the design process, the authors propose Scene-GRPO, a critic-guided framework built on GRPO (Generative Reward Policy Optimization). This approach uses quality scores to steer the model toward better design decisions while reducing the number of useless iterations and error corrections.

Results: 89% Accuracy and Beating GPT-5.2

Experiments showed Scene-Designer's dual edge — both in the quality of the generated 3D scenes and in its ability to critique them:

Pairwise scene comparison — Scene-Critic achieved 89% accuracy at picking the better of two scenes

Four-scene ranking — the model showed 76% accuracy at ordering four variants by quality

Outperforming GPT-5.2 — the resulting metrics beat this baseline model

These results show that the framework doesn't just generate strong 3D scenes — it has genuinely developed the ability to assess them, the very design mindset needed for quality work. The combination of two specialized datasets (35,000 design examples and 15,000 critique examples) alongside the critic-guided Scene-GRPO approach produced a model capable of outperforming prior solutions in this space.