/ZeroNVS

Primary LanguagePythonApache License 2.0Apache-2.0

ZeroNVS

This is the offical code release for ZeroNVS: Zero-Shot 360-Degree View Synthesis from a Single Real Image.

teaser image

What is in this repository: 3D SDS distillation code, evaluation code, trained models

In this repository, we currently provide code to reproduce our main evaluations and also to run ZeroNVS to distill NeRFs from your own images. This includes scripts to reproduce the main metrics on DTU and Mip-NeRF 360 datasets.

How do I train my own diffusion models?

Check out the companion repository, https://github.com/kylesargent/zeronvs_diffusion.

Acknowledgement

This codebase is heavily built off existing codebases for 3D-aware diffusion model training and 3D SDS distillation, namely Zero-1-to-3 and threestudio. If you use ZeroNVS, please consider also citing these great contributions.

Requirements

The code has been tested on an A100 GPU with 40GB of memory.

To get the code:

git clone https://github.com/kylesargent/zeronvs.git
cd zeronvs

To set up the environment, use the following sequence of commands. The exact setup that will work for you might be platform dependent. Note: it's normal for installing tiny-cuda-nn to take a long time.

conda create -n zeronvs python=3.8 pip
conda activate zeronvs

pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit
pip install ninja git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

pip install -r requirements-zeronvs.txt
pip install nerfacc -f https://nerfacc-bucket.s3.us-west-2.amazonaws.com/whl/torch-2.0.0_cu118.html

Finally, be sure to initialize and pull the code in the zeronvs_diffusion submodule.

cd zeronvs_diffusion
git submodule init
git submodule update

cd zero123
pip install -e .
cd ..

cd ..

Data and models

Since we have experimented with a variety of datasets in ZeroNVS, the codebase consumes a few different types of data formats.

To download all the relevant data and models, you can run the following commands within the zeronvs conda environment

gdown --fuzzy https://drive.google.com/file/d/1q0oMpp2Vy09-0LA-JXpo_ZoX2PH5j8oP/view?usp=sharing
gdown --fuzzy https://drive.google.com/file/d/1aTSmJa8Oo2qCc2Ce2kT90MHEA6UTSBKj/view?usp=drive_link
gdown --fuzzy https://drive.google.com/file/d/17WEMfs2HABJcdf4JmuIM3ti0uz37lSZg/view?usp=sharing

unzip dtu_dataset.zip

MipNeRF360 dataset

You can download it here. Be sure to set the appropriate path in resources.py

DTU dataset

Download it here (hosted by the PixelNeRF authors). Be sure to unzip it and then set the approriate path in resources.py

Your own images

Store them as 256x256 png images and pass them to launch_inference.sh (details below).

Models

We release our main model, trained with our $\mathbf{M}_{\mathrm{6DoF+1,~viewer}}$ parameterization on CO3D, RealEstate10K, and ACID. You can download it here. We use this one model for all our main results.

Inference

Evaluation is performed by distilling a NeRF for each of the scenes in the dataset. DTU has 15 scenes and the Mip-NeRF 360 dataset has 7 scenes. Since NeRF distillation takes ~3 hours, running the full eval can take quite some time, especially if you only have 1 GPU.

Note that you can still achieve good performance with much faster config options; for instance, reduced resolution, batch size, number of training steps, or some combination. The code as-is is just intended to reproduce the results from the paper.

After downloading the data and models, you can run the evals via either launch_eval_dtu.sh or launch_eval_mipnerf360. The metrics for each scene will be saved in metrics.json files which you must average to get the final performance.

We provide the expected performance for individual scenes in the tables below. Note that there is some randomness inherent in SDS distillation, so you may not get exactly these numbers (though the performance should be quite close, especially on average).

DTU (expected performance)

ssim psnr lpips scene_uid manual_gt_to_pred_scale
0.6094 13.2329 0.2988 8.0 1.2
0.1739 8.4278 0.5783 21.0 1.4
0.6311 14.1864 0.2332 30.0 1.5
0.2992 8.9569 0.5117 31.0 1.4
0.3862 14.049 0.3611 34.0 1.4
0.3495 12.6771 0.4659 38.0 1.3
0.4612 12.2447 0.3729 40.0 1.2
0.4657 12.5998 0.3794 41.0 1.3
0.369 11.241 0.4441 45.0 1.4
0.4456 17.0177 0.4322 55.0 1.2
0.5724 12.6056 0.2639 63.0 1.5
0.5384 12.1564 0.2725 82.0 1.5
0.5434 16.0902 0.3811 103.0 1.5
0.6353 19.5588 0.349 110.0 1.3
0.5529 18.2336 0.3613 114.0 1.3

Mip-NeRF 360 (expected performance)

ssim psnr lpips scene_uid manual_gt_to_pred_scale
0.1707 13.184 0.6536 bicycle 1.0
0.3164 13.1137 0.6122 bonsai 1.0
0.2473 12.2189 0.6823 counter 0.9
0.207 15.2817 0.5366 garden 1.0
0.254 13.2983 0.6245 kitchen 0.9
0.3431 11.8591 0.5928 room 0.9
0.1396 13.124 0.6717 stump 1.1

Running on your own images

Use the script launch_inference.sh. You will need to specify the image path, field-of-view, camera elevation, and content scale. These don't need to be exact, but badly wrong values will cause convergence failure.

Citation

If you use ZeroNVS, please cite via:

@article{zeronvs,
    author = {Sargent, Kyle and Li, Zizhang and Shah, Tanmay and Herrmann, Charles and Yu, Hong-Xing and Zhang, Yunzhi and Chan, Eric Ryan and Lagun, Dmitry and Fei-Fei, Li and Sun, Deqing and Wu, Jiajun},       
    title = {{ZeroNVS}: Zero-Shot 360-Degree View Synthesis from a Single Real Image},
    journal={arXiv preprint arXiv:2310.17994},
    year={2023}
}