This repository contains the official implementation for DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation.
accelerate.mp4
- 2023.12.22: add experimental support for ImageDream, check imagedream.yaml.
- 2023.12.14: add support for Stable-Zero123, check image_sai.yaml.
- 2023.10.21: add support for MVDream, check text_mv.yaml.
pip install -r requirements.txt
# a modified gaussian splatting (+ depth, alpha rendering)
git clone --recursive https://github.com/ashawkey/diff-gaussian-rasterization
pip install ./diff-gaussian-rasterization
# simple-knn
pip install ./simple-knn
# nvdiffrast
pip install git+https://github.com/NVlabs/nvdiffrast/
# kiuikit
pip install git+https://github.com/ashawkey/kiuikit
# To use MVdream, also install:
pip install git+https://github.com/bytedance/MVDream
# To use ImageDream, also install:
pip install git+https://github.com/bytedance/ImageDream/#subdirectory=extern/ImageDream
Tested on:
- Ubuntu 22 with torch 1.12 & CUDA 11.6 on a V100.
- Windows 10 with torch 2.1 & CUDA 12.1 on a 3070.
Image-to-3D:
### preprocess
# background removal and recentering, save rgba at 256x256
python process.py data/name.jpg
# save at a larger resolution
python process.py data/name.jpg --size 512
# process all jpg images under a dir
python process.py data
### training gaussian stage
# train 500 iters (~1min) and export ckpt & coarse_mesh to logs
python main.py --config configs/image.yaml input=data/name_rgba.png save_path=name
# gui mode (supports visualizing training)
python main.py --config configs/image.yaml input=data/name_rgba.png save_path=name gui=True
# load and visualize a saved ckpt
python main.py --config configs/image.yaml load=logs/name_model.ply gui=True
# use an estimated elevation angle if image is not front-view (e.g., common looking-down image can use -30)
python main.py --config configs/image.yaml input=data/name_rgba.png save_path=name elevation=-30
### training mesh stage
# auto load coarse_mesh and refine 50 iters (~1min), export fine_mesh to logs
python main2.py --config configs/image.yaml input=data/name_rgba.png save_path=name
# specify coarse mesh path explicity
python main2.py --config configs/image.yaml input=data/name_rgba.png save_path=name mesh=logs/name_mesh.obj
# gui mode
python main2.py --config configs/image.yaml input=data/name_rgba.png save_path=name gui=True
# export glb instead of obj
python main2.py --config configs/image.yaml input=data/name_rgba.png save_path=name mesh_format=glb
### visualization
# gui for visualizing mesh
# `kire` is short for `python -m kiui.render`
kire logs/name.obj
# save 360 degree video of mesh (can run without gui)
kire logs/name.obj --save_video name.mp4 --wogui
# save 8 view images of mesh (can run without gui)
kire logs/name.obj --save images/name/ --wogui
### evaluation of CLIP-similarity
python -m kiui.cli.clip_sim data/name_rgba.png logs/name.obj
python -m kiui.cli.clip_sim_text prompt logs/name.obj
Please check ./configs/image.yaml
for more options.
Image-to-3D (stable-zero123):
### training gaussian stage
python main.py --config configs/image_sai.yaml input=data/name_rgba.png save_path=name
### training mesh stage
python main2.py --config configs/image_sai.yaml input=data/name_rgba.png save_path=name
Text-to-3D:
### training gaussian stage
python main.py --config configs/text.yaml prompt="a photo of an icecream" save_path=icecream
### training mesh stage
python main2.py --config configs/text.yaml prompt="a photo of an icecream" save_path=icecream
Please check ./configs/text.yaml
for more options.
Text-to-3D (MVDream):
### training gaussian stage
python main.py --config configs/text_mv.yaml prompt="a plush toy of a corgi nurse" save_path=corgi_nurse
### training mesh stage
python main2.py --config configs/text_mv.yaml prompt="a plush toy of a corgi nurse" save_path=corgi_nurse
Please check ./configs/text_mv.yaml
for more options.
Image+Text-to-3D (ImageDream):
### training gaussian stage
python main.py --config configs/imagedream.yaml input=data/ghost_rgba.png prompt="a ghost eating hamburger" save_path=ghost
### training mesh stage
python main2.py --config configs/imagedream.yaml input=data/ghost_rgba.png prompt="a ghost eating hamburger" save_path=ghost
Helper scripts:
# run all image samples (*_rgba.png) in ./data
python scripts/runall.py --dir ./data --gpu 0
# run all text samples (hardcoded in runall_sd.py)
python scripts/runall_sd.py --gpu 0
# export all ./logs/*.obj to mp4 in ./videos
python scripts/convert_obj_to_video.py --dir ./logs
Gradio Demo:
python gradio_app.py
- The world & camera coordinate system is the same as OpenGL:
World Camera
+y up target
| | /
| | /
|______+x |/______right
/ /
/ /
/ /
+z forward
elevation: in (-90, 90), from +y to -y is (-90, 90)
azimuth: in (-180, 180), from +z to +x is (0, 90)
- Trouble shooting OpenGL errors (e.g.,
[F glutil.cpp:338] eglInitialize() failed
):
# either try to install OpenGL correctly (usually installed with the Nvidia driver), or use force_cuda_rast:
python main.py --config configs/image_sai.yaml input=data/name_rgba.png save_path=name force_cuda_rast=True
kire mesh.obj --force_cuda_rast
#Bugs
- Link not found: microsoft/DeepSpeed#3410
- Tinycuda: https://github.com/NVlabs/tiny-cuda-nn
- Pymeshlab: 2022.post3 + 3DTopia/LGM#2
This work is built on many amazing research works and open-source projects, thanks a lot to all the authors for sharing!
@article{tang2023dreamgaussian,
title={DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation},
author={Tang, Jiaxiang and Ren, Jiawei and Zhou, Hang and Liu, Ziwei and Zeng, Gang},
journal={arXiv preprint arXiv:2309.16653},
year={2023}
}