ReViV: Reconstructing the Viewer and the View in 4D from Monocular Egocentric Video

ECCV 2026 · Code · Paper · Website

ReViV reconstructs egocentric 4D from a single monocular RGB video in one feed-forward pass. It recovers the viewer, meaning the body and hand motion of the camera wearer, and the view, meaning depth, camera trajectory and gaze.

Checkpoints

This repository holds two self-contained inference sets, one per subfolder. Each set contains the main model (reviv_main.pth), the five detokenizers (reviv_tok_{body,cam,gaze,lhand,rhand}.pth) and a norm_stats/ folder with the matching mean/std files. Use the detokenizers and norm stats from the same set as the main model.

Set	Input	Predicts
metric_depth/	32-frame 512×512 clip @ 16 fps (Cosmos DV8x16x16)	metric 512×512 depth, camera, gaze, body, hands
reviv_500b/	16-frame 256×256 clip @ 8 fps (Cosmos DV4x8x8) + raw clip	relative 256×256 depth, camera, gaze, body, hands

Depth is decoded with the NVIDIA Cosmos tokenizer, which you download separately (see the GitHub README).

Usage

The ReViV environment (environment.yaml) pins huggingface_hub==0.25.2, which provides huggingface-cli:

huggingface-cli download SeanLyu/reviv --include "metric_depth/*" --local-dir reviv_checkpoints
export REVIV_CKPT_ROOT=reviv_checkpoints/metric_depth

Do not upgrade huggingface_hub inside the ReViV environment, because diffusers 0.20 requires a version below 0.26. Outside that environment, a newer huggingface_hub offers hf download with the same arguments.

Then run the demos as described in the GitHub repository. To use the 256 pathway, replace metric_depth with reviv_500b.

License

The model weights are released under the Sample Code License (see LICENSE), which permits non-commercial use only. The code on GitHub is released under Apache 2.0.

The weights were derived from the following datasets: Ego-Exo4D, HoloAssist, HOT3D (Aria and Quest), ARCTIC, TACO, H2O, EgoGen and Nymeria. Depth supervision consists of pseudo-labels generated with Video Depth Anything. This repository does not redistribute any dataset or derived annotation. Each dataset and model listed above has its own license and terms of use, which apply in addition to the license granted here.

Use of the model weights is limited to non-commercial research purposes.

Citation

@inproceedings{lyu2026reviv,
  title     = {ReViV: Reconstructing the Viewer and the View in 4D from Monocular Egocentric Video},
  author    = {Lyu, Xiaozhong and Li, Gen and Qian, Zhiyin and Zhang, Xucong and Pollefeys, Marc and Tang, Siyu},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}