Instead of producing images/ videos from scene descriptions, extract scene descriptions from simulated images using inverse generative AI

https://techxplore.com/news/2025-08-ai-method-reconstructs-3d-scene.html#google_vignette

"generalizes well across new images or different scenarios/ extracted patterns traditionally difficult to interpret... gradually adjusts model's internal parameters... solve vision tasks, such as tracking, as test-time optimization problems... new datasets not needed for training on... compares rendered image with real observed image, backpropagating difference through differentiable rendering function/ 3D generation model to update its inputs... provides 3D explanations of its perceived world... robotics, autonomous driving"