Stable3R

NRGBD geometric context probe: StreamVGGT mean ATE changes substantially as intermediate frames are added, while Stable3R remains lower and more stable. — Context sensitivity on NRGBD.

Stable geometric
references.

Refine historical memory and reduce first-frame bias—without revising past predictions.

Abstract

Causal attention and KV-caching enable streaming 3D reconstruction, but information-poor early caches and first-frame anchoring can introduce persistent geometric bias. Stable3R addresses both sources of instability: lookahead attention refines historical caches as new observations arrive, while a frame-equivariant architecture with prefix-relative pose supervision reduces reliance on a single initial frame. Together, these designs improve reconstruction quality and camera pose estimation while preserving a strictly causal inference schedule.

Method

Lookahead attention

New observations refine historical memory.

Prefix-relative supervision

Learn relative poses across the observed prefix.

Parallel training

Original paper diagram of block-wise causal attention, diagonal dependency order, update kernels, and attention kernels.

Interactive comparison

Scene

Common-input reconstruction comparison

Point error0 cm≥10 cm

StreamVGGT

Ours

Ground truth

RGB-D reference

Selected scene only

Loading real point clouds…

Details & sources

Matched RGB frames · 518 × 392 · BF16 · full history. Same camera, point size and error scale; no model-specific filtering. RGB-D ground truth is used for evaluation. Error colors measure distance to the full GT cloud, not the display subset.

Evaluation: first-camera GT coordinates, median depth shift/scale normalization, 224 × 224 center crop, correspondence Sim(3), and point-to-point ICP (0.1 m). Acc ↓: predicted-to-GT distance; Comp ↓: GT-to-predicted distance; NC ↑: normal consistency.

Drag to rotate, Shift/right-drag to pan, scroll/pinch to zoom. Arrow keys rotate; +/− zoom; R resets. Views are synchronized; there is no automatic rotation.

Three selected examples, not a dataset-wide average. Frame IDs, metrics & source hashes

Paper results

Original reconstruction comparison of CUT3R, TTT3R, StreamVGGT, LongStream, and Stable3R, with staircase and dynamic indoor examples under sparse and mid-length settings. — Sparse (top) and 25-frame (bottom) reconstructions. Stable3R: rightmost.

3D reconstruction

Dataset

NRGBD 25-frame reconstruction results from the manuscript
Method	Accuracy ↓		Completeness ↓		Normal consistency ↑
Method	Mean	Median	Mean	Median	Mean	Median
CUT3R	0.080	0.046	0.053	0.025	0.820	0.958
TTT3R	0.087	0.049	0.055	0.028	0.808	0.948
StreamVGGT	0.079	0.055	0.059	0.032	0.846	0.983
LongStream	0.078	0.050	0.052	0.024	0.836	0.975
Stable3R	0.022	0.014	0.020	0.010	0.884	0.987

Protocol & metrics

Paper Table 1. Uniformly sampled 25-frame sequences.

Accuracy: predicted-to-reference distance. Completeness: reference-to-predicted distance. Normal consistency: surface-normal agreement. ↓ Lower is better; ↑ higher is better.

Bold: best; underline: second best. Paper table data

Camera pose estimation