As-Plausible-As-Possible (APAP)

Project Page | Paper | arXiv

Seungwoo Yoo*¹, Kunho Kim*¹, Vladimir G. Kim², Minhyuk Sung¹ (* co-first authors)

¹KAIST, ²Adobe Research

This is the reference implementation of As-Plausible-As-Possible: Plausibility-Aware Mesh Deformation Using 2D Diffusion Priors (CVPR 2024).

Get Started

Clone the repository and create a Python environment:

git clone https://github.com/KAIST-Visual-AI-Group/APAP
cd APAP
conda create --name apap python=3.9
conda install pytorch==2.0.1 torchvision==0.15.2 torchaudio==2.0.2 pytorch-cuda=11.7 -c pytorch -c nvidia
conda install pytorch-sparse -c pyg
pip install wandb
pip install diffusers==0.19.0
pip install accelerate transformers ninja
pip install cholespy libigl
pip install imageio[ffmpeg] jaxtyping tyro
pip install fpsample trimesh pymeshlab pyrender

You also need to follow the instructions in the following repositories to install dependencies:

• nvdiffrast
• Semantic-SAM

We provide data necessary to run our codes via Google Drive. Specifically, you can download

• APAP-Bench 2D
• APAP-Bench 3D
• LoRA checkpoints

Download the files from the link and place them under the directory data. After that, the directory structure should look like:

APAP
├── data
│   ├── apap_2d  # APAP-Bench 2D
│   ├── apap_3d  # APAP-Bench 3D
│   ├── lora_ckpts  # LoRA checkpoint files
│   ├── pretrained_models  # Additional pretrained models (e.g., SAM)
│   └── ...
├── ext
├── scripts
├── src
├── environment.yaml
└── README.md

Making Deformations using APAP-Bench

To run 3D mesh deformation experiments using APAP-Bench (3D), run:

python scripts/exp/batch/batch_deform_meshes.py \
--data-list-path configs/deform_meshes/data/apap_3d.txt
--out-root outputs/apap-3d
--gpu-ids 0

Note that the experiments can be parallelized by specifying multiple GPU IDs via the argument --gpu-ids.

Similarly, 2D mesh deformation experiments using APAP-Bench (2D) can be done by running:

python scripts/exp/batch/batch_deform_meshes.py \
--data-list-path configs/deform_meshes/data/apap_2d-all.txt
--out-root outputs/apap-2d
--gpu-ids 0

Fine-tuning Stable Diffusion using LoRA

We directly adapt Dreambooth training script from diffusers without modification. For convenience, we provide a batch script that allows users to train multiple LoRAs in parallel. To run the script, simply execute:

python scripts/lora/batch_train_dreambooth_lora.py \
--data-list-path configs/lora_train/apap_3d.txt \
--exp-group-name apap-3d-lora \
--out-root outputs/lora_ckpts/apap-3d \
--gpu-ids 0

This will produce LoRA checkpoints, each fine-tuned to the renderings of meshes in APAP-Bench (3D). Note that each row of a training config file consists of two items - object_name and data_dir. The object_name is used to automatically populate a text prompt used during fine-tuning and the data_dir is a directory containing images for fine-tuning.
After training, the outputs are arranged into a directory structure as follows:

{out-root}
├── object_name1
│   ├── 0000  # Identifier for image dataset 
│   └── ...
├── object_name2
│   ├── 0000  # Identifier for image dataset 
│   └── ...
├── object_name3
│   ├── 0000  # Identifier for image dataset 
│   └── ...
└── ...

The checkpoint directories can be passed to the script deform_meshes.py via the command-line argument --lora-dir.

Citation

Please consider citing our work if you find this codebase useful:

@inproceedings{yoo2024apap,
  title = {{As-Plausible-As-Possible: Plausibility-Aware Mesh Deformation Using 2D Diffusion Priors}},
  author = {Yoo, Seungwoo and Kim, Kunho and Kim, Vladimir G. and Sung, Minhyuk},
  booktitle = {CVPR},
  year = {2024},
}