论文概要
Research Area: Computer Vision Authors: Stefano Esposito, Naama Pearl, Polina Karpikova, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Andreas Geiger, Jonathon Luiten Published: 2026-10-08 arXiv: 2610.12388
Summary
Reconstructing complete 3D object assets from monocular or sparse multi-view observations remains challenging. Generative 3D foundation models can complete object geometry beyond the observed views, but their predictions may not faithfully reproduce the observed geometry, appearance, or pose.
GenIA is a framework for test-time input-aligned generation that grounds SAM3D's generative prior in geometric and photometric observations without retraining the foundation model.
Key technical contributions
- Pose improvement: Translation and scale are derived from geometry while retaining the learned rotation prior.
- Appearance alignment via:
- visibility-biased attention,
- cross-observation fusion,
- differentiable rendering guidance during denoising.
- Optional post-denoising refinement further adapts the appearance latents, lightweight decoder adapters, and object placement.
Dynamic objects
The framework also supports externally provided geometry: given the temporal shape of a dynamic object, GenIA recovers a shared, input-aligned canonical appearance and stable world-space placement.
Results
On both synthetic and real benchmarks, GenIA outperforms recent optimization-based per-frame image-to-3D and video-to-4D methods in object pose prediction and reconstruction across monocular, multi-view, and dynamic inputs.
---
*Auto-collected on 2026-10-12*