Google Research Benchmarks Gemini's 3D Object Generation Through Code

GoogleGoogle

Google Research introduced 3DCodeBench, a new benchmark evaluating AI models' ability to generate 3D objects using code. This benchmark, presented at CVPR2026, demonstrates how agentic AI can autonomously create complex 3D assets, highlighting the role of iterative refinement in improving model performance.

Google Research unveiled 3DCodeBench, a benchmark for evaluating AI models in agentic procedural 3D modeling via code. It assesses how well models, including Gemini variants, generate diverse 3D objects by writing and executing Blender Python scripts. The benchmark includes 212 object categories and evaluates 12 frontier VLMs (vision-language models processing text and images).
Models Evaluated
12 frontier VLMs
Object Categories
212
3D Objects with Code
13,000
Multi-view Renders
52,000
Ranking System
Human-preference Elo
Top Performing Model (Elo)
GPT-5.5 (1167)

3DCodeBench reveals that while models can generate code, they often lack physical-world understanding, producing disconnected or misaligned structures. However, multi-turn refinement with deterministic feedback significantly improves performance, showing that the agentic harness is as crucial as the model itself. This emphasizes the importance of iterative processes in complex AI tasks.

The benchmark's 3DCodeArena provides human-preference Elo rankings for evaluated models, allowing comparison of performance and cost across paid frontier models. 3DCodeBench also offers open data for training and evaluation, including Blender code, multi-view renders, and LLM-generated captions, supporting further research and development in 3D content creation.

3DCodeBench framework evaluates VLM performance in generating procedural Python code for 3D object creation and rendering.
View the full update on 3dcodebench.com
Google Research
Google Research
@GoogleResearch
X

See Project Astra 3D at the #CVPR2026 Google booth #557 at 5:30pm! The team presents 3DCodeBench, demonstrating Gemini models' proficiency in generating diverse 3D objects through code execution. Presented by Lei Shu & Yipeng Gao. https://t.co/XlZKytbZwE @GoogleDeepMind https://t.co/OmZ0bz4zVG

11retweets99likes
View on X

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update