Generative Multiplane Images (GMPI)

Making a 2D GAN 3D-Aware

ECCV 2022 (Oral)

Generative Multiplane Images: Making a 2D GAN 3D-Aware, ECCV 2022 (Oral).
Xiaoming Zhao, Fangchang Ma, David Güera Cobo, Zhile Ren, Alexander G. Schwing, and Alex Colburn.

Project Page | Paper

Table of Contents

• Environment Setup
• Using Pretrained Checkpoints
  • Render a Single Image
  • Render a Video
  • Extract a Mesh
• Training and Evaluation
• Citation
• License

Environment Setup

This code has been tested on Ubuntu 18.04 with CUDA 10.2.

conda env create -f environment.yml

Using Pretrained Checkpoints

Download Checkpoints

cd /path/to/this/repo
export GMPI_ROOT=$PWD

Please download our pretrained checkpoints from the release page or this link and place them under ${GMPI_ROOT}/ckpts. The structure should be:

.
+-- ckpts
|  +-- gmpi_pretrained
|  |  +-- FFHQ256
|  |  +-- FFHQ512
|  |  +-- FFHQ1024
|  |  +-- AFHQCat
|  |  +-- MetFaces

We use the following variables for illustration purposes.

# This can be FFHQ256, FFHQ512, FFHQ1024, AFHQCat, or MetFaces
export DATASET_NAME=FFHQ1024
export OUTPUT_DIR=${GMPI_ROOT}/ckpts/gmpi_pretrained/${DATASET_NAME}

# Set this to your favourate seed
export SEED=589

# - When psi = 1.0 there is no truncation, which is used for quantitative results in the paper.
# - To obtain better qualitative results, use psi < 1.0.
export TRUNCATION_PSI=1.0

Render a Single Image

The following command renders an image ${OUTPUT_DIR}/rendered.png, along with:

• mpi_alpha.png: alpha maps for all planes,
• mpi_rgb.png: the same RGB texture for all planes,
• mpi_rgba.png: RGB-alpha images for all planes.

conda activate gmpi && \
export PYTHONPATH=${GMPI_ROOT}:${GMPI_ROOT}/gmpi/models:$PYTHONPATH && \
python ${GMPI_ROOT}/gmpi/eval/vis/render_video.py \
--ckpt_path ${OUTPUT_DIR}/generator.pth \
--output_dir ${OUTPUT_DIR} \
--seeds ${SEED} \
--nplanes 96 \
--truncation_psi ${TRUNCATION_PSI} \
--exp_config ${OUTPUT_DIR}/config.pth \
--render_single_image 1

Note: We use nplanes = 96 in the paper for reporting quantitative and qualitative results, but GMPI is able to produce high-quality results even with 32 planes. Use a small nplanes (e.g., 32) if your run into CUDA out-of-memoory errors.

Render a Video

The following command renders a video in ${OUTPUT_DIR}, along with:

• video_rgb.mp4: video for the RGB rendering,
• video_depth.mp4: video for the depth rendering.

conda activate gmpi && \
export PYTHONPATH=${GMPI_ROOT}:${GMPI_ROOT}/gmpi/models:$PYTHONPATH && \
python ${GMPI_ROOT}/gmpi/eval/vis/render_video.py \
--ckpt_path ${OUTPUT_DIR}/generator.pth \
--output_dir ${OUTPUT_DIR} \
--seeds ${SEED} \
--nplanes 96 \
--truncation_psi ${TRUNCATION_PSI} \
--exp_config ${OUTPUT_DIR}/config.pth \
--render_single_image 0 \
--horizontal_cam_move 1

Notes:

• Set nplanes to some small number (e.g., 32) if your run into CUDA out-of-memoory errors.
• Set horizontal_cam_move to 0 if you want a video with vertical camera motion.

Extract a Mesh

The following command produces a mesh ${OUTPUT_DIR}/mesh_${TRUNCATION_PSI}.ply.

conda activate gmpi && \
export PYTHONPATH=${GMPI_ROOT}:${GMPI_ROOT}/gmpi/models:$PYTHONPATH && \
python ${GMPI_ROOT}/gmpi/eval/vis/extract_mesh.py \
--ckpt_path ${OUTPUT_DIR}/generator.pth \
--dataset ${DATASET_NAME} \
--save_dir ${OUTPUT_DIR} \
--exp_config ${OUTPUT_DIR}/config.pth \
--stylegan2_sanity_check 0 \
--truncation_psi ${TRUNCATION_PSI} \
--seed ${SEED} \
--chunk_n_planes -1

Notes:

• Set chunk_n_planes to some small positive numbers (e.g., 64) if your run into CUDA out-of-memoory errors.
• It may take several minutes to extract a mesh with Marching Cubes, for resolutions of 512 or above.

Training and Evaluation

Please refer to TRAIN_EVAL.md for more details.

Citation

Xiaoming Zhao, Fangchang Ma, David Güera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative Multiplane Images: Making a 2D GAN 3D-Aware. ECCV 2022.

@inproceedings{zhao-gmpi2022,
  title = {Generative Multiplane Images: Making a 2D GAN 3D-Aware},
  author = {Xiaoming Zhao and Fangchang Ma and David Güera and Zhile Ren and Alexander G. Schwing and Alex Colburn},
  booktitle = {Proc. ECCV},
  year = {2022},
}

License

This sample code is released under the LICENSE terms.

Disclaimer

Some of this software was built on Nvidia codebase, as noted within the applicable files, and such Nvidia code is available under its own terms at https://github.com/NVlabs/stylegan2-ada-pytorch. The authors of this software are not responsible for the contents of third-party websites.