AI creates 3D scenes from a single photo, but cannot see its own mistakes

Researchers have developed a method called LEGO-Anything. It converts a single photo into code that can build a 3D scene in the program Blender. Blender is a widely used program for creating 3D graphics, and the scene can be edited afterward.

The method also comes with a test, a so-called benchmark, which measures how well different AI agents handle the task. An AI agent is an AI system that carries out tasks step by step on its own. GPT-6 Astra performs best, with up to 53 percent accuracy in recreating the scene.

Cannot judge its own work

The biggest weakness applies to all the agents tested: they are no better than a coin toss at assessing whether their own 3D model is geometrically correct. Geometric accuracy is about whether shapes, sizes and positions in the 3D scene match the image.

That means the agents can deliver a scene that looks finished without knowing themselves whether it is right. If you are going to use such a tool, it is therefore still necessary for a human to check the result.