license: apache-2.0
library_name: pytorch
pipeline_tag: image-to-3d
tags:
- 4d-reconstruction
- dynamic-scene
- scene-flow
- point-cloud
- point-tracking
- query-based
UniQuery4R: Unified 4D Scene Reconstruction from a Single Query
Paper | Project Page | Code
Tiancheng Chen1, Sheng Tang1, Wenhua Jin1,2, Weiqi Zhang3, Juntong Fang3, Junsheng Zhou3, Zesong Li1
1Kosmo Research 2Automotive Engineering Department, Jilin University 3School of Software, Tsinghua University
UniQuery4R is a query-conditioned feed-forward framework for unified 4D scene reconstruction from a single query. A variable-length multi-view clip is jointly encoded once; at decoding time, a continuous source-pixel query q = (u, v, src, tgt) selects the source and target views via source-to-target cross-attention. Each query jointly predicts target correspondence (warp2d), target-time 3D position (warp3d), scene flow (warp3d_delta, parameterized as direction x magnitude) and source depth, while camera parameters are estimated per view.
Model files
| File | Description |
|---|---|
uniquery4r.pth |
UniQuery4R weights (ViT-g encoder, ~1.65B parameters) |
Usage
Install the code from the GitHub repository, then either download this file manually or let huggingface_hub fetch it automatically by passing the repo id:
python scripts/infer_4d.py --checkpoint Kosmo-Research/UniQuery4R --input path/to/images --output outputs/demo
from uniquery4r.models.uniquery4r import UniQuery4R
model = UniQuery4R.from_pretrained("Kosmo-Research/UniQuery4R", device="cuda")
import torch
images = ... # [S, 3, H, W] float tensor in [0, 1]
with torch.inference_mode():
predictions = model(images)
See the GitHub README for the full CLI, motion-mask, viewer and WorldTrack-evaluation workflows.
Citation
@article{chen2026uniquery4r,
title={{UniQuery4R}: Unified {4D} Scene Reconstruction from a Single Query},
author={Chen, Tiancheng and Tang, Sheng and Jin, Wenhua and Zhang, Weiqi and Fang, Juntong and Zhou, Junsheng and Li, Zesong},
journal={arXiv preprint arXiv:2608.17283},
year={2026}
}
License
Released under the Apache License 2.0. Portions of the codebase are adapted from VGGT and remain under the VGGT License (non-commercial); see NOTICE for details.