/MVDream-threestudio

3D generation code for MVDream

Primary LanguagePythonApache License 2.0Apache-2.0

MVDream - threestudio

Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, Xiao Yang

| Project Page | Paper | Gallery | Comparison |

  • This code is forked from threestudio for SDS and 3D Generation using MVDream.
  • For diffusion model and 2D image generation, check original MVDream repo.

mvdream-threestudio-teaser

Installation

Install threestudio

This part is the same as original threestudio. Skip it if you already have installed the environment.

See installation.md for additional information, including installation via Docker.

  • You must have an NVIDIA graphics card with at least 20GB VRAM and have CUDA installed.
  • Install Python >= 3.8.
  • (Optional, Recommended) Create a virtual environment:
python3 -m virtualenv venv
. venv/bin/activate

# Newer pip versions, e.g. pip-23.x, can be much faster than old versions, e.g. pip-20.x.
# For instance, it caches the wheels of git packages to avoid unnecessarily rebuilding them later.
python3 -m pip install --upgrade pip
  • Install PyTorch >= 1.12. We have tested on torch1.12.1+cu113 and torch2.0.0+cu118, but other versions should also work fine.
# torch1.12.1+cu113
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113
# or torch2.0.0+cu118
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118
  • (Optional, Recommended) Install ninja to speed up the compilation of CUDA extensions:
pip install ninja
  • Install dependencies:
pip install -r requirements.txt

Install MVDream

MVDream multi-view diffusion model is provided in a different codebase. Install it by:

git clone https://github.com/bytedance/MVDream extern/MVDream
pip install -e extern/MVDream 

Quickstart

We currently provide two configurations for MVDream, one without soft-shading and one with it. The one without shading is more effecient in both memory and time. You can run it by

# MVDream without shading (memory efficient)
python launch.py --config configs/mvdream-sd21.yaml --train --gpu 0 system.prompt_processor.prompt="an astronaut riding a horse"

In the paper, we use the configuration with soft-shading. It would need an A100 GPU in most cases to compute normal:

# MVDream with shading (used in paper)
python launch.py --config configs/mvdream-sd21-shading.yaml --train --gpu 0 system.prompt_processor.prompt="an astronaut riding a horse"

Resume from checkpoints

If you want to resume from a checkpoint, do:

# resume training from the last checkpoint, you may replace last.ckpt with any other checkpoints
python launch.py --config path/to/trial/dir/configs/parsed.yaml --train --gpu 0 resume=path/to/trial/dir/ckpts/last.ckpt
# if the training has completed, you can still continue training for a longer time by setting trainer.max_steps
python launch.py --config path/to/trial/dir/configs/parsed.yaml --train --gpu 0 resume=path/to/trial/dir/ckpts/last.ckpt trainer.max_steps=20000
# you can also perform testing using resumed checkpoints
python launch.py --config path/to/trial/dir/configs/parsed.yaml --test --gpu 0 resume=path/to/trial/dir/ckpts/last.ckpt
# note that the above commands use parsed configuration files from previous trials
# which will continue using the same trial directory
# if you want to save to a new trial directory, replace parsed.yaml with raw.yaml in the command

# only load weights from saved checkpoint but dont resume training (i.e. dont load optimizer state):
python launch.py --config path/to/trial/dir/configs/parsed.yaml --train --gpu 0 system.weights=path/to/trial/dir/ckpts/last.ckpt

Tips

  • Preview. Generating 3D content with SDS would a take a lot of time. So we suggest to use the 2D multi-view image generation MVDream to test if the model can really understand the text before using it for 3D generation.
  • Rescale Factor. We introducte rescale adjustment from Shanchuan et al. to alleviate the texture over-saturation from large CFG guidance. However, in some cases, we find it to cause floating noises in the generated scene and consequently OOM issue. Therefore we reduce the rescale factor from 0.7 in original paper to 0.5. However, if you still encounter such a problem, please try to further reduce system.guidance.recon_std_rescale=0.3.

Credits

This code is built on the threestudio-project. Thanks to the maintainers for their contribution to the community!

Citing

If you find MVDream helpful, please consider citing:

@article{shi2023MVDream,
  author = {Shi, Yichun and Wang, Peng and Ye, Jianglong and Mai, Long and Li, Kejie and Yang, Xiao},
  title = {MVDream: Multi-view Diffusion for 3D Generation},
  journal = {arXiv:2308.16512},
  year = {2023},
}