Infer3D: Inverse Generative Modeling Enables Robust 3D Reconstruction

Infer3D: Inverse Generative Modeling Enables Robust 3D Reconstruction


Authors: Amani Kiruga, Mark Hamilton, Joshua B. Tenenbaum, William T. Freeman, Vincent Sitzmann, Yilun Du, Ayush Tewari


Subjects: Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Robotics (cs.RO)


arXiv: 2611.25061 [cs.CV] — Submitted 18 November 2026


Project page: amanikiruga.github.io/infer3d-site




Overview


Humans can understand the 3D structure of scenes even from highly unfamiliar or extreme viewpoints. Cognitive science suggests this generalization stems from a top-down generative inference strategy. In contrast, state-of-the-art 3D reconstruction neural networks typically rely on amortized, feedforward inference. While fast, this pattern-matching approach degrades when faced with out-of-distribution (OOD) scene conditions unseen during training.


Inspired by the generative nature of human vision, we propose Infer3D, an inverse generative modeling framework that recovers the latent variables of a generative model to best match the monocular input.


Method


To prevent degenerate flat outputs, Infer3D couples the generative image prior with a frozen 3D lifting network that constrains the optimization space to a valid, multi-view-consistent 3D manifold.


Key Contributions


  • Inverse generative modeling for 3D reconstruction: Rather than learning a direct feedforward mapping from images to 3D, Infer3D performs inference over the latent space of a generative model, aligning the rendered output with the input image.
  • Manifold-constrained optimization: A frozen 3D lifting network anchors the optimization to a physically valid, multi-view-consistent 3D manifold, avoiding degenerate flat reconstructions.
  • Adaptive test-time compute: Mirroring how human perception requires more time to process confusing stimuli, Infer3D introduces an adaptive inference strategy that expends more test-time compute on ambiguous OOD examples while efficiently routing in-distribution inputs.

Results


Across a variety of controlled synthetic and real-world tasks, Infer3D demonstrates superior OOD reconstruction accuracy compared to feedforward networks trained on equal or larger datasets.


Significance in 2026


As 3D reconstruction models are increasingly deployed in robotics, embodied AI, and mixed-reality pipelines, robustness to unseen viewpoints and scene conditions has become a central bottleneck. Infer3D's shift from amortized feedforward inference toward generative, optimization-based inference—paired with adaptive test-time compute—reflects a broader 2026 trend of trading inference-time computation for generalization, echoing advances in reasoning and test-time scaling seen across the field.




23 pages, 12 figures, 15 tables. Includes references and appendices.

via ArXiv CV

Related