LEGO-Anything turns photos into editable Blender scenes, but AI agents still can't judge their own 3D accuracy

A new method called LEGO-Anything can convert a single photograph into editable Blender code for a 3D scene, according to research highlighted by The Decoder. The pipeline is designed to let AI agents generate structured, modifiable geometry rather than a static mesh, opening the door to iterative editing by text prompt. But a companion benchmark reveals a stubborn problem: the same agents that build the scenes cannot reliably tell whether their reconstructions are geometrically correct.

The benchmark tested several leading AI agents on 3D reconstruction tasks. GPT-6 Astra finished first with up to 53 percent accuracy, meaning even the strongest system got roughly half of the geometric details wrong by the benchmark's own measure. That figure is not a rounding error; it is the current ceiling for single-image 3D scene generation when the output must be editable code rather than a visual approximation.

More striking is the self-evaluation gap. Across all tested agents, the ability to judge their own geometric accuracy was no better than a coin flip. In other words, an agent can produce a Blender scene that looks plausible from the original camera angle while containing wrong dimensions, misplaced objects, or impossible geometry, and it has no internal signal to flag the error. That makes autonomous iteration risky, because the agent may 'fix' a correct part and leave a broken one untouched.

The LEGO-Anything approach matters because editable Blender code is far more useful than a fixed 3D model for game development, robotics simulation, and visual effects. Artists and engineers can adjust objects, materials, and lighting after generation. But without a reliable self-check, human review remains mandatory at every step, which limits how much of the workflow can be automated.

The results point to a broader issue in agent research: generation quality is improving faster than verification quality. GPT-6 Astra's 53 percent accuracy is a benchmark high, yet the coin-flip self-assessment means the model cannot be trusted to know when it has succeeded. Closing that gap, not just raising reconstruction accuracy, is the next milestone for anyone building agents that operate in 3D space.