Code for How to Use Diffusion Priors under Sparse Views? (NeurIPS 2024)

Installation

Ubuntu 22.04, CUDA 11.3, PyTorch 1.12.1

conda env create --file environment.yaml
conda activate ipsm

pip install ./submodules/diff-gaussian-rasterization-confidence ./simple-knn

Pre-trained Models Preparation

mkdir pretrained_models
cd pretrained_models

Download StableDiffusion-v1.5, StableDiffusionInpainting-v1.5, MiDaS, BLIP to ./pretrained_models/. (NOTE: Stable Diffusion V1.5 and Stable Diffusion Inpainting V1.5 cannot be downloaded from the original repo, but the same weight can be obtained from other clone repo.)

Data Preparation

LLFF

1. Download LLFF from the official download link.

2. Run COLMAP to obtain initial point clouds with sparse views:

   python tools/colmap_llff.py
   

3. Randomly select one image from sparse views and run BLIP to obtain its blip-based text results:

   python ./scripts/script_for_blip.py
   

4. The data format is supposed to be:

   |- <scene>
       |- 3_views
       |- images
       |- images_4
       |- images_8
       |- sparse
       |- blip_rst.txt
       |- poses_bounds.npy
       |- ...
   

DTU

1. Download DTU dataset

   • Download the DTU dataset "Rectified (123 GB)" from the official website, and extract it.
   • Download masks (used for evaluation only) from this link.

2. Preprocess following DNGaussian

   • Poses: following gaussian-splatting, run convert.py to get the poses and the undistorted images by COLMAP.
   • Render Path: following LLFF to get the poses_bounds.npy from the COLMAP data. (Optional)

3. Run COLMAP to obtain initial point clouds with sparse views:

   python tools/colmap_dtu.py
   

4. Randomly select one image from sparse views and run BLIP to obtain its blip-based text results:

   python blip_script.py
   

5. The data format is supposed to be:

   |- <scene>
       |- 3_views
       |- images
       |- images_2
       |- images_4
       |- images_8
       |- mask
       |- sparse
       |- blip_rst.txt
       |- poses_bounds.npy
       |- ...
   

Training & Rendering & Evaluating

Train & Render & Evaluate IPSM-Gaussian on the LLFF dataset with 3 views:

python ./scripts/script_for_llff.py

Train & Render & Evaluate IPSM-Gaussian on the LLFF dataset with 3 views:

python ./scripts/script_for_dtu.py

Acknowledgement

This code is developed on gaussian-splatting, FSGS, and DNGaussian. Thanks for these great projects!

Citation

@inproceedings{
wang2024how,
title={How to Use Diffusion Priors under Sparse Views?},
author={Qisen Wang and Yifan Zhao and Jiawei Ma and Jia Li},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
url={https://openreview.net/forum?id=i6BBclCymR}
}