AI Agent 从照片构建 3D 场景但难自评
Image-to-Code 让编码代理迭代生成 Blender 场景代码。
为什么重要
为单图 3D 重建代理提供了可执行输出和自动化评测框架。
阅读原文the-decoder.com
AI 解读（根据原文生成）
University of Maryland 和 AWS 研究者提出 LEGO-Anything：编码代理接收单张图像，为 Blender 编写并迭代修正代码，输出可检查、编辑和查询的 3D 场景程序。配套 LEGO-Bench 用专业模拟器场景渲染的图像提供隐藏几何真值，评测可用性、重建准确度和外观匹配。原文未说明代码或数据是否开放。
要点（AI 据原文整理，数字均已与原文核对）
- LEGO-Bench contains 208 images from 104 indoor and outdoor scenes and uses 443 registered assets.
- The benchmark scores each scene on three axes: validity, reconstruction, and appearance.
- All six tested GPT configurations delivered a working scene almost every time.
- GPT-6 Astra hit 53.4 percent on indoor scenes and 39.6 percent on outdoor scenes.
- Weaker configurations scored around 15 percent.
- Astra's score on an office test subset jumped from 32.3 to 61.8 percent when the reasoning budget increased.
