Point a camera at a room. Get back a complete, editable 3D model of everything in it: the sofa, the table, the chairs, each one a clean mesh you can rotate, move, and drop into a game engine or a design tool. That is PixARMesh, a single-image 3D scene reconstruction method from UC San Diego and Lambda, accepted at CVPR 2026.
The problem: one photo, a whole scene
A single photo shows one viewpoint. Half the scene is occluded, depth is ambiguous, and the objects overlap. Reconstructing accurate 3D geometry from that is a long-standing, fundamentally underdetermined problem.
Prior single-view 2D-to-3D scene reconstruction methods predominantly rely on implicit representations (e.g., occupancy fields or signed distance functions) or multi-stage pipelines that separately estimate scene layout, object geometry, and poses. These approaches often require computationally expensive post-processing, accumulate errors across stages, and produce meshes that are not directly suitable for downstream graphics or simulation applications.
Who needs this
This sits upstream of a lot of work. A robot needs a 3D map of its environment before it can plan or act. A self-driving stack reconstructs 3D from camera images before control happens. A game developer or an interior designer wants a real 3D asset from a single reference photo, not hours of manual modeling.
How PixARMesh works
PixARMesh borrows a trick from language models. Tokenize everything, the image, the object pose, the mesh, into one sequence, then let an autoregressive Transformer predict it token by token. One forward pass produces both where each object sits (pose) and its shape (mesh). No signed distance fields. No separate layout-optimization loop.
The key novelty is bringing autoregressive, mesh-native generation to single-image scene-level reconstruction, enabling direct generation of artist-ready meshes without relying on implicit representations or multi-stage pipelines. Pixel-aligned image features combined with global scene context through cross-attention improve reconstruction quality, particularly under occlusion.
The results
On the 3D-FRONT benchmark, PixARMesh sets a new state of the art at the scene level: scene-level F-Score of 33.55% vs. 25.00% for the prior best (DepR), and Chamfer Distance cut from 0.153 to 0.099. It does this while producing compact, artist-ready meshes (a few thousand faces per object) instead of dense, over-tessellated surfaces.
High-quality 3D scene reconstruction is becoming a core workload for industries such as robotics, autonomous systems, AR/VR, digital twins, gaming, and embodied AI. All of these require significant GPU resources for training and inference. More efficient and end-to-end reconstruction methods reduce computational overhead while enabling customers to build and deploy large-scale 3D AI applications on cloud infrastructure. In turn, supporting these emerging workloads strengthens Lambda's AI platform and attracts developers building next-generation multimodal and spatial AI systems.
Explore the interactive demo: https://mlpc-ucsd.github.io/PixARMesh/
Credits: UC San Diego + Lambda. Authors: Xiang Zhang, Sohyun Yoo, Hongrui Wu, Chuan Li, Jianwen Xie, Zhuowen Tu. CVPR 2026.