News

• 2026.05: Added minimal DL3DV training and evaluation code.
• 2025.12: Initial release.

Quick Start

git clone --recursive https://github.com/QitaoZhao/E-RayZer.git
cd E-RayZer

bash scripts/setup_train_env.sh
source .venv/bin/activate

If Python 3.10 is available under another name, pass it explicitly:

PYTHON_BIN=/path/to/python3.10 bash scripts/setup_train_env.sh

If you cloned without --recursive, initialize the submodule first:

git submodule update --init --recursive

Checkpoints

The Gradio demo automatically downloads missing weights on first launch. You can also download them manually from Hugging Face and place them under checkpoints/:

• checkpoints/erayzer_multi.pt: multi-dataset model, used by default.
• checkpoints/erayzer_dl3dv.pt: DL3DV-only model.

Pass --ckpt /path/to/checkpoint.pt if you keep weights elsewhere.

Demo

python gradio_app.py \
	--config config/erayzer_inference.yaml \
	--device cuda:0 \
	--share

Upload multi-view RGB images, or use one of the bundled examples. The demo writes predicted camera poses, target renders, point_cloud.glb, render_video.mp4, and a zip archive under experiments/inference/erayzer.

Training and Evaluation

We provide a minimal DL3DV training and evaluation release:

• Data preparation
• Training
• Evaluation

The code here is a re-implementation and differs from the original version developed at Adobe.

Repository Map

• gradio_app.py: End-to-end Gradio demo UI.
• app_core/engine.py: Inference wrapper for configs, checkpoints, renders, Gaussian point clouds, and videos.
• erayzer_core/: E-RayZer model, transformer blocks, losses, camera utilities, and Gaussian renderer.
• config/erayzer_inference.yaml: Default inference configuration.
• config/erayzer_train_dl3dv.yaml: Minimal DL3DV training configuration.
• config/erayzer_eval_dl3dv.yaml: Minimal DL3DV evaluation configuration.
• training/: Local DL3DV dataset loader and dataset factory.
• scripts/: Environment setup, DL3DV data conversion, manifest checking, training, and evaluation scripts.
• examples/: Curated multi-view examples for quick inference checks.
• third_party/gsplat/: Differentiable Gaussian splatting ops with intrinsics gradient support.

Citation

If you use E-RayZer in academic or industrial research, please cite:

@inproceedings{zhao2026erayzer,
  title     = {E-RayZer: Self-supervised 3D Reconstruction as Spatial Visual Pre-training},
  author    = {Qitao Zhao and Hao Tan and Qianqian Wang and Sai Bi and Kai Zhang and Kalyan Sunkavalli and Shubham Tulsiani and Hanwen Jiang},
  booktitle = {CVPR},
  year      = {2026}
}

Related Project

This project is inspired by and builds upon RayZer. We strongly encourage readers to check it out if you have not already.

License

• Code: MIT License, see LICENSE.
• Model weights: Adobe Research License, see LICENSE-WEIGHTS. The model weights are not covered by the MIT License.

Acknowledgements

This work was partially done at Adobe Research, where Qitao Zhao worked as a Research Scientist Intern. We thank Zhengqi Li for insightful advice. We also thank Frederic Fortier-Chouinard, Jiashun Wang, Yanbo Xu, Zihan Wang, and members of the Physical Perception Lab for helpful discussions.