3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

RELATION-CENTRIC PARAMETER-EFFICIENT ADAPTATION

3R-Adapter

Retrieval, Rewiring, and Refinement for Efficient Adaptation of 3D Reconstruction Model

Zixiang Ai · Zhenyu Cui · Xu Zou · Jiahuan Zhou · NeurIPS 2026 Spotlight

Recover cross-view relations under domain shift while preserving pretrained geometric priors.

Figure 3 from the paper: the complete 3R-Adapter pipeline
Figure 3 · Full method pipeline
TITLE / FIGURE 301 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

01 / MOTIVATION · SCALING BOUNDARY

Pretraining scales with data.
Long-tail supervision does not.

Feed-forward 3D reconstruction models benefit from large-scale pretraining, but that recipe assumes abundant and diverse geometric supervision.

Figure 1 crop: scaling of pretrained foundation models and the long-tail data bottleneck
Figure 1(a) · The 3D data-scaling bottleneck
01

What scales

More pretrained images and 3D supervision improve the learned geometry and cross-view correspondence prior.

02

Where scaling stops

Dense 3D ground truth is expensive. Rare conditions and structured downstream tasks remain sparsely represented.

03

The resulting gap

A strong frozen model still encounters target distributions that are not covered by the head of the pretraining data.

Therefore: long-tail adaptation cannot be reduced to collecting another web-scale supervised 3D dataset.
INTRODUCTION / FIGURE 1(A)02 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

02 / MOTIVATION · DISTRIBUTION SHIFT

The head and the tail are not
the same visual distribution.

Daytime imagery represents a data-rich, in-distribution regime. Extreme low-light imagery changes visibility, texture evidence, and cross-view matching reliability.

HEAD / DATA-RICHDaytime
Figure 2 crop: standard daytime scene and concentrated cross-view attention
Observed in the pretrained regime
DOMAIN
SHIFT→

Appearance statistics change while the reconstruction task still requires the same underlying geometry.

TAIL / DATA-SPARSEExtreme low light
Figure 2 crop: nighttime scene and diffuse cross-view attention
Long-tail target distribution
Low lightSevere appearance degradation
Human–sceneNon-rigid motion and cross-entity relations
Test-time adaptationScene-specific shift without ground-truth geometry
INTRODUCTION / SECTION 2 / FIGURE 203 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

03 / OBSERVATION AND INSIGHT

What changes is not only the image.
The relational route also collapses.

We trace the same selected query across reference views. The comparison reveals how the model's cross-view routing changes under the head-to-tail distribution shift.

Figure 2: concentrated cross-view attention in daytime and diffuse attention in nighttime
Figure 2 · Cross-view attention visualization using the same query-view logic
DAYTIMEConcentrated attention

The query attends to spatially consistent physical counterparts across reference views.

→
NIGHTTIMEDiffuse and fragmented attention

The same relational routing becomes less informative and correspondences become unreliable.

OBSERVED INSIGHTRelational attention structure collapse

The adaptation bottleneck lies partly in unstable cross-view relations, not only weakened local features.

SECTION 2 / FIGURE 204 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

04 / MOTIVATION · DESIGN INSIGHT

If relations collapse,
feature modulation alone is indirect.

Conventional PEFT adjusts features and relies on attention to reorganize implicitly. We instead place trainable capacity directly at the relational bottleneck.

Figure 1 crop: conventional PEFT versus 3R-Adapter relational adaptation
Figure 1(b) · From implicit feature adjustment to explicit relational adaptation
OBSERVATION

Relational routes become unstable

Cross-view attention is no longer concentrated on reliable counterparts.

↓
REFORMULATION

Adaptation as relation recovery

Retrieve task-relevant candidates, explicitly rewire associations, and inject them without overwriting the pretrained prior.

↓
METHOD

Retrieval · Rewiring · Refinement

A lightweight trainable branch attached to a frozen reconstruction backbone.

INTRODUCTION / FIGURE 1(B)05 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

03 / METHOD OVERVIEW

One relation-centric pipeline, three cohesive operations

The pretrained reconstruction backbone is frozen. 3R-Adapter is inserted as a lightweight trainable branch that explicitly adapts cross-view relations.

Figure 3 from the paper: Deformable Retrieval, Relational Rewiring, and Consistency Refinement
Figure 3 · Overall pipeline of 3R-Adapter
01Deformable Retrieval

Predict sparse candidate locations and routing weights.

02Relational Rewiring

Materialize explicit cross-view associations by sampling.

03Consistency Refinement

Inject the rewired relation through a controlled residual update.

SECTION 4 / FIGURE 306 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

04 / METHOD · RETRIEVAL

Deformable Retrieval

Input: self-view query tokens and neighboring reference tokens.
Output: low-rank features, K offsets, and K routing weights per query.

Original Figure 3 crop: Deformable Retrieval
Figure 3 · Deformable Retrieval module
01

Select reference views

For target view t, the sliding window provides a coarse locality prior over the view-level relation graph.

N(t) = {v ∣ |v − t| ≤ w}   O(N²) → O(Nw)
02

Project both token streams

The Self-View Project and Reference-View Project map C-dimensional backbone features into a shared r-dimensional adapter space, with r ≪ C.

x̃tu = ϕself(xtu),   X̃v = ϕref(Xv)
03

Predict candidate offsets and weights

Two heads act on each compact self-view query. The Offset Head predicts where to sample; the Weight Head predicts each candidate's contribution.

{Δp̂ku}k=1…K = ψΔ(x̃tu),   {wku}k=1…K = ψw(x̃tu)
Interface to Rewiring: sparse candidate offsets and weights, plus projected reference grids. Sampling happens in the next stage.
SECTION 4 / FIGURE 3 / ORIGINAL MODULE CROP07 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

05 / METHOD · REWIRING

Relational Rewiring

Input: predicted offsets, routing weights, and projected reference features.
Output: a relation-enhanced low-rank representation for each self-view query.

Original Figure 3 crop: Relational Rewiring
Figure 3 · Relational Rewiring module
01

Bound the spatial search

Apply a bounded transform to raw offsets, then shift the query's spatial coordinate. The radius ρ limits the deformable search range.

Δpku = ρ tanh(Δp̂ku),   pku = pu + Δpku
02

Sample each reference feature field

Reshape X̃ᵛ from HW × r to H × W × r. Grid interpolation evaluates the reference feature field at each candidate coordinate.

fv,ku = B(X̃v, pku)   B: bilinear sampling
03

Aggregate sparse cross-view candidates

Sum sampled features over selected reference views and K candidates using the predicted weights. This materializes the sparse relation graph.

rtu = ∑v∈N(t) ∑k=1K wku fv,ku
Interface to Refinement: rtu ∈ ℝr carries the sampled cross-view evidence. Default K = 4 and ρ = 0.20.
SECTION 4 / FIGURE 3 / ORIGINAL MODULE CROP08 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

06 / METHOD · REFINEMENT

Consistency Refinement

Input: the self-view latent token and its rewired relational representation.
Output: a residual-corrected token in the original backbone feature space.

Original Figure 3 crop: Consistency Refinement
Figure 3 · Consistency Refinement module
01

Blend relational information

The α-blending block combines the original low-rank query with the aggregated cross-view evidence. α controls the strength of relation injection.

htu = (1 − α)x̃tu + αrtu
02

Project the fused representation upward

The Refinement Project maps the fused r-dimensional representation back to the C-dimensional backbone space to obtain a feature correction.

Δxtu = ϕup(htu) ∈ ℝC
03

Apply a scaled residual update

Keep the original feature on the main stream and add the relation correction. γ balances adaptation strength against prior preservation.

x̂tu = xtu + γ Δxtu
Output to the frozen prediction stream: corrected tokens support downstream geometry. Defaults: α = 0.10 and γ = 0.30.
SECTION 4 / FIGURE 3 / ORIGINAL MODULE CROP09 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

05 / EXPERIMENTS

Three adaptation settings probe different relational failures

All pretrained backbones are frozen. Baselines follow matched training protocols and evaluation settings.

LOW-LIGHT

Oxford Day&Night

Camera pose estimation and multi-view scene reconstruction under severe appearance degradation.

CUT3R · π³
TEST-TIME ADAPTATION

7-Scenes · NRGBD

Two to five sparse views; independent per-scene optimization with self-supervised pointmap consistency.

DUSt3R
HUMAN–SCENE

BEDLAM · EMDB

Joint scene geometry and SMPL reconstruction with non-rigid motion and cross-entity interactions.

CUT3R
OPTIMIZATIONAdamW · 8× RTX 5090

Only lightweight adapter parameters are updated.

SCENE METRICSAcc. · Comp. · Chamfer · Normal

Distance metrics are reported in meters.

POSE METRICSATE · RPE-T · RPE-R

Translation in meters; rotation in degrees.

HUMAN METRICSPA/Metric-PVE · PA/Metric-MPJPE

Human errors are reported in millimeters.

Task objectives: Llow=Lgeo+λposeLpose · Lhs=Lscene+λmeshLmesh+λjointLjoint+λsmplLsmpl · Ltta=Σ‖P(r1)−P(r2)‖1
SECTION 5 / APPENDIX A.3–A.610 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

06 / LOW-LIGHT RECONSTRUCTION

Strongest PEFT performance on both CUT3R and π³

Oxford Day&Night · camera pose estimation and scene reconstruction. Lower is better except Normal consistency.

2.8Mtrainable parameters
on CUT3R
0.581Chamfer on CUT3R
vs. 0.791 NightAdapter
0.633Chamfer on π³
vs. 0.936 backbone

On CUT3R, Chamfer is 26.5% lower than the strongest PEFT baseline. On π³, it is 32.4% lower than the original backbone.

Table 1 from the paper: Oxford Day and Night comparison
Table 1 · Full quantitative comparison
TABLE 1 / OXFORD DAY&NIGHT11 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

07 / CROSS-TASK ADAPTATION

The relation-centric design transfers beyond low light

Test-time adaptation and joint human-scene reconstruction share the need for stable cross-view or cross-entity relations.

Test-time adaptation

Table 2 from the paper: test-time adaptation on 7-Scenes and NRGBD
Table 2 · 7-Scenes and NRGBD
2.5Mtrainable parameters0.216 / 0.205Chamfer on 7-Scenes / NRGBD

Best PEFT results on both datasets; 8.9% and 6.8% lower Chamfer than T-LoRA.

Joint human–scene reconstruction

Table 3 from the paper: BEDLAM human-scene reconstruction
Table 3 · BEDLAM validation set
Figure 4 from the paper: qualitative human-scene reconstruction results
Figure 4 · Qualitative results

Strongest overall PEFT results with the fewest PEFT parameters; competitive with task-specific architectures.

TABLES 2–3 / FIGURE 412 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

08 / ABLATION AND SENSITIVITY

Rewiring provides the key structural gain; refinement completes it

Oxford Day&Night · CUT3R backbone · Chamfer Distance ↓

Table 4 from the paper: component ablation
Table 4 · Component ablation
Figure 5 from the paper: hyperparameter ablation for K, rho, alpha, and gamma
Figure 5 · Hyperparameter sensitivity
Retrieval only1.201 → 1.152

Sparse querying reduces noisy interactions, but candidate locations alone are insufficient.

+ Relational Rewiring1.152 → 0.842

Explicit reorganization produces the largest structural improvement.

Full 3R-Adapter0.581

The three modules are complementary; Normal consistency reaches 0.550.

Default settingK=4 · ρ=.20 · α=.10 · γ=.30

Excessive search or relational injection introduces redundancy or disturbs pretrained features.

TABLE 4 / FIGURE 513 / 14
3R ADAPTER / RESEARCH TALKNEURIPS 2026 SPOTLIGHT

CONCLUSION

Adapt relational structure,
not only features.

3R-Adapter places lightweight trainable capacity at the relational bottleneck of feed-forward 3D reconstruction.

Figure 3 from the paper: 3R-Adapter pipeline
Retrieval · Rewiring · Refinement
01
Failure mode

Long-tail domain shift is associated with relational attention structure collapse.

02
Method

Sparse retrieval, explicit rewiring, and residual refinement recover cross-view associations.

03
Evidence

Consistent PEFT gains appear in low-light, test-time, and human-scene adaptation.

Limitations

The method depends on useful backbone features and geometric priors. A fixed sliding window may restrict long-range relation modeling in very long sequences or under large viewpoint changes.

Future: adaptive view selection and hierarchical relational memory.
CONCLUSION / APPENDIX A.214 / 14
CLICK OR ESC TO CLOSE
← → Navigate   F Fullscreen   N Notes   O Overview   P Auto-play   Click figures to zoom