Implementation for paper G. Li, Q. Fen, S. Qian, and et al. High Fidelity GAN Inversion via Prior Multi-Subspace Feature Composition, accepted for Association for Advancements of Artificial Intelligence (AAAI) 2021. For more details, please refer to my personal website or read the PDF.
- /models: This directory contains most of the SOTA generators, e.g., PGGAN, StyleGAN2 and BigGAN. The original repo is from mGANprior. Please download necessary checkpoints from the following links and place them in the /models/pretrain/tensorflow, the project will convert tensorflow-based checkpoints to pytorch version, and store them in /models/pretrain/pytorch. For a complete link, please refer to /models/README.MD.
| PGGAN Official | |||
|---|---|---|---|
| Face | |||
| celebahq-1024x1024 | |||
| Indoor Scene | |||
| bedroom-256x256 | livingroom-256x256 | diningroom-256x256 | kitchen-256x256 |
| Outdoor Scene | |||
| churchoutdoor-256x256 | tower-256x256 | bridge-256x256 | |
| Animal | |||
| person-256x256 | cat-256x256 | dog-256x256 | bird-256x256 |
| horse-256x256 | sheep-256x256 | cow-256x256 |
-
/bin: contains some testing generated images and pretrained affinity matrixs.
-
/examples: contains some target images.
-
Note: The original repo has some mistakes in loading StyleGAN, and I have corrected them. I also add many personal utilities to the repo, e.g., getting intermediate features for calculating subspaces. It could be messy, so you can pull requests if you are interested.
Please use following command to install the environment.
python -m pip install -r requirements.txt
I also provide the conda environment in file conda_env.yaml if you are interested.
conda env create -f conda_env.yaml
Take CelebaHQ as example, you can find useful subspaces by running the command.
python discover_subspace.py \
--outputs=./TRAIN --inversion_type=PGGAN-Layerwise \
--gan_model=pggan_celebahq --layer=3 --iterations=200 \
--optimization=Adam --lr=0.0001 --report_image=5 \
--report_model=5 --batch_size=4 --exp_id=SelfExpress \
--init_type=Zeros --beta=1 --cluster_numbers=6 --alpha=0.2 \
--subspace_dimension=6 --sparse_type=L1 --power=2.0
Place the pretrained affinity matrix to args.matrix_dir and run the following command.
python PmSFC.py --outputs=./TRAIN \
--inversion_type=PGGAN-Layerwise --gan_model=pggan_celebahq \
--layer=3 --iterations=20000 --optimization=Adam --lr=0.0001 \
--report_image=5 --report_model=5 --batch_size=4 --exp_id=SelfExpress \
--beta0=1 --beta1=1 --n_subs=6 --alpha=0.2 --d_subs=6 \
--sparse_type=L1 --power=2.0 \
--matrix_dir=./bin/pggan/celebahq/matrix/layer3/value50.pkl
You should get similar results like this:
