Meta Reality Labs and University of Tübingen · Input-Aligned 3D Reconstruction

GenIA

A method, posted 8 October (arXiv 2610.12388), that makes Meta's frozen SAM 3D model match real photos and video at generation time without retraining. It takes object translation and scale from observed geometry while keeping SAM 3D's learned rotation estimate, and aligns appearance with visibility-biased attention, fusion across views and differentiable-rendering guidance during denoising; an optional last step fine-tunes small adapters. It handles single images, multiple views and dynamic input, and on dynamic sequences costs about 13 times SAM 3D's single-image runtime against about 160 times for competing methods, author-reported. Andreas Geiger is among the authors. Builds on SAM 3D (week 47 of 2025).