Latent Bridge

Latent BridgeFrom Video Latents to 3D Geometry
via Feature Alignment

Shuai Zhang1, Zhanpeng Ouyang2,‡, Zhengqing Chen2, Ruohong Mei2, Shaoyu Chen2,
Weiqiang Ren2, Hui Sun2, Qian Zhang2, Wenyu Liu1, Xinggang Wang1,†

1 Huazhong University of Science and Technology2 Horizon Robotics

‡ Project leader   ·   † Corresponding author

PaperCode

Method Overview

Alignment training · Generative inference
Paper Figure 2. Alignment training with frozen video encoder and DINOv3 teacher; generative inference through Latent Bridge into VGGT-Omega; encoder-specific Wan2.1 and V-RAE Bridge architectures.
Figure 2. Only the Bridge is trained. A shared video latent supports both RGB decoding and 3D geometry.View original PDF ↗Explore in 3D ↓
THE IDEA

Recover geometry
before decoding color.

Video generation and 3D reconstruction speak different representational languages. Connecting them usually means decoding RGB frames, then encoding those frames all over again.

Latent Bridge learns the connection instead. A representation-specific network maps Wan2.1 or V-RAE latents into the native feature space of frozen VGGT-Ω. The same latent can then produce both a video and cameras, depth, and 3D point clouds.

Only the Bridge is trainedNo GT geometry supervisionFrozen geometry model

Qualitative Results

01 / 41

3D point clouds

Drag any view to control all five.
01 / 41
All five full-density viewers load when you scroll here.
Original five-method comparison ↗

Applications to other I2V and T2V models.

01 / 49The viewer loads when this section is visible.

Quantitative Results

Geometry results from Table 1 of the paper

THE MATCHED-LATENT COMPARISON

Less decoding.
More direct.

Same generated latents. Same frozen Ω aggregator and heads. Only the route into the geometry model changes.

Table 3 in the paper ↗
Wan2.1 1.61×
RGB route
7.92 s
Bridge
4.92 s
V-RAE 1.24×
RGB route
5.76 s
Bridge
4.64 s

Citation

arXiv:2610.28521 · 2026
@misc{zhang2026latentbridge,
  title={Latent Bridge: From Video Latents to 3D Geometry via Feature Alignment},
  author={Shuai Zhang and Zhanpeng Ouyang and Zhengqing Chen and Ruohong Mei and Shaoyu Chen and Weiqiang Ren and Hui Sun and Qian Zhang and Wenyu Liu and Xinggang Wang},
  year={2026},
  eprint={2610.28521},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.28521}
}