Breaking

Tencent Hunyuan Team Releases WorldStereo for 3D Point Clouds From One Image

July 3, 2026 at 03:33 EDT

Media Generation
Research & Papers
Open Source

Researchers from Tencent Hunyuan and Zhejiang University have released WorldStereo, a framework that generates multi-view consistent videos and 3D point clouds from a single reference image. Published as a CVPR 2026 accepted paper, its code and model weights are now available.

CVPR 2026 · Tencent Hunyuan × Zhejiang University

WorldStereo: One Image → Consistent Video + Dense 3D Point Clouds

A single framework that bridges camera-guided video generation and 3D scene reconstruction — producing multi-view consistent video and dense point clouds at once via 3D geometric memories.

reference image in — video + point cloud out

4-step

fast inference via DMD, no joint training

geometric memory modules: GGM + SSM

The pipeline

single image + camera.json → memories → outputs

Single image + camera

extrinsic/intrinsic, panorama or multi-trajectory

→

Geometric memories

GGM injects coarse structure · SSM preserves detail

→

Video + point clouds

.ply + depth, normals, poses, Gaussian Splatting

Benchmark scores — WorldStereo 2.0

Two families of metrics: lower is better for camera-pose errors, higher is better for quality & reconstruction.

Camera pose error ↓

0.492

RotErr

0.968

TransErr

1.768

ATE

Reconstruction ↑

51.27

MipNeRF360
F1

65.79

MipNeRF360
AUC

4.205

Q-Align visual quality

89.43

CLIP-I consistency

Why it stands out

Memory across multiple trajectories, not one-pass or autoregressive generation
Precise camera control plus multi-trajectory consistency
Beats SEVA, Gen3C, Lyra, FlashWorld and Dust3R-style methods

Still to verify

Speed-versus-quality trade-off of the DMD fast version
Detail preservation in complex scenes
Broad real-world use limited — code released only recently

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.