The feed-forward pipeline lifts visual features into 3D, predicts instance structure, and uses canonicalized object evidence to condition shape and material generation. A highly compressed VAE supports batched processing of objects, after which meshes are placed back into the estimated metric scene layout.
Fire3D is designed for scene reconstruction and simulation-data workflows that need actual object meshes. The project reports reconstruction in under a minute and releases inference code, training code, models, and example inputs. Its reference protocols use substantial GPU memory, so local deployment requires checking the documented environment.
