Abundant images do not imply abundant 3D labels.
Pre-training can cover common visual conditions at scale, yet high-quality camera, depth, point-map, and human geometry supervision remains costly to acquire.
3R-Adapter: Retrieval, Rewiring, and Refinement for Efficient Adaptation of 3D Reconstruction Model
NeurIPS 2026 Spotlight
1Peking University 2Tsinghua University 3Huazhong University of Science and Technology
TL;DR
Long-tail domain shifts break cross-view correspondences in feed-forward 3D reconstruction, a failure that channel-wise PEFT does not directly address. 3R-Adapter freezes the backbone and restores relations through Deformable Retrieval, Relational Rewiring, and Consistency Refinement. This relation-centric design improves low-light reconstruction, test-time adaptation, and joint human-scene reconstruction with lightweight trainable modules.
Two-minute overview
3R-Adapter in two minutes.
An overview of the motivation, three-stage method, and reconstruction results.
Motivation
Scaling stops where 3D supervision becomes scarce.
Feed-forward reconstruction models learn geometry from large-scale pre-training, but dense 3D ground truth is expensive. The remaining data are not merely fewer samples. They occupy difficult, shifted regimes where the model's learned relational routes no longer hold.
Head data are daytime; the tail changes the evidence.
Low light suppresses texture, dynamic humans break static-scene assumptions, and test-time adaptation offers only a few unlabeled views from one scene.
The relationship between views collapses first.
Under shift, attention becomes diffuse and fragmented. Features may remain plausible locally while cross-view correspondence becomes unreliable.
Observation to insight
Adapt relations, not only channels.
Conventional adapters and low-rank reparameterization adjust features and leave relational recovery implicit. Our evidence points to a more direct objective: reconstruct sparse, task-adaptive cross-view associations.
Relational attention structure collapse is the shared failure mode across low-light reconstruction, joint human-scene reconstruction, and test-time adaptation. 3R-Adapter addresses this structure explicitly while preserving the backbone's pre-trained prior.
Method
Three operations recover a stable relational route.
The backbone remains frozen. A lightweight branch searches a bounded neighborhood, materializes sparse cross-view associations, and injects the recovered relation through a gated residual update.
R1 / Retrieve
Deformable Retrieval
Project tokens into a compact latent space, predict a small set of bounded offsets and routing weights, and restrict reference views with a sliding window. This changes view interactions from all-pairs to a structured local search.
R2 / Rewire
Relational Rewiring
Use bilinear sampling at the predicted locations, then aggregate candidates with learned routing weights. The result is an explicit, sparse relational graph instead of unconstrained all-to-all attention.
R3 / Refine
Consistency Refinement
Fuse the self-view latent token with its rewired relation, project back to the backbone dimension, and apply a learned residual gate. The update improves consistency without overwriting the pre-trained representation.
Experiments
One relation-centric adapter, three long-tail settings.
We evaluate low-light reconstruction on Oxford Day&Night, self-supervised test-time adaptation on 7-Scenes and NRGBD, and joint human-scene reconstruction on BEDLAM.
Oxford Day&Night
Pi3 backbone
Pi3 backbone
0.633
Chamfer distance, 3.2M parameters
Oxford Day&Night
CUT3R backbone
CUT3R backbone
0.581
Chamfer distance, 2.8M parameters
Test-time adaptation
7-Scenes / NRGBD
7-Scenes / NRGBD
.216 / .205
Chamfer distance, 2.5M parameters
BEDLAM
Human3R backbone
Human3R backbone
44.4
PA-PVE in mm, 36.1M parameters
Citation
Build on 3R-Adapter.
@inproceedings{ai2026threeRAdapter,
title = {3R-Adapter: Retrieval, Rewiring, and Refinement
for Efficient Adaptation of 3D Reconstruction Model},
author = {Ai, Zixiang and Cui, Zhenyu and Zou, Xu and Zhou, Jiahuan},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026}
}