LEGO-Anything turns a single photo into 3D scenes, but agents can't self-assess geometry

On October 3, 2026, The Decoder reported on a new approach called LEGO-Anything that converts a single photo into editable Blender code for 3D scenes. The report says the method comes with a companion benchmark, and that GPT-6 Astra leads it with reconstruction accuracy of up to 53 percent.

The same coverage points to a shortcoming shared by every agent tested: none of them can tell whether the geometry they reconstructed is actually correct. Asked to rate their own geometric accuracy, the agents did no better than a coin flip. In practice, that means they can turn a photo into 3D scene code, but they have no dependable sense of whether the output holds up geometrically, a gap the article frames as the central finding.

The report does not say who developed LEGO-Anything, how it was released, or what the benchmark tasks look like in detail, and it offers no follow-up work. As of that report, the state of things is this: the photo-to-Blender-code pipeline works, GPT-6 Astra holds the top spot on the accompanying benchmark at up to 53 percent reconstruction accuracy, and self-assessment of geometric accuracy remains unsolved across all the agents that were tested.