BlobGAN: Spatially Disentangled Scene Representations
_{Official PyTorch Implementation}

Paper | Project Page | Video | Interactive Demo

BlobGAN.mp4

This repository contains:

• 🚂 Pre-trained BlobGAN models on three datasets: bedrooms, conference rooms, and a combination of kitchens, living rooms, and dining rooms
• 💻 Code based on PyTorch Lightning ⚡ and Hydra 🐍 which fully supports CPU, single GPU, or multi GPU/node training and inference

We also provide an 📓 interactive demo notebook to help get started using our model. Download this notebook and run it on your own Python environment, or test it out on Colab. You can:

• 🖌️️ Generate and edit realistic images with an interactive UI
• 📹 Create animated videos showing off your edited scenes
• 📸 (new!) Upload your own image and convert it into blobs!

And, coming soon:

• 🧬 More edits, as shown in the paper! Code for cloning, restyling, rotating, and reshaping blobs.

Setup

Run the commands below one at a time to download the latest version of the BlobGAN code, create a Conda environment, and install necessary packages and utilities.

git clone https://github.com/dave-epstein/blobgan.git
mkdir -p blobgan/logs/wandb
conda create -y -n blobgan python=3.9
conda activate blobgan
conda install -y pytorch=1.11.0 torchvision=0.12.0 torchaudio cudatoolkit=11.3 -c pytorch
conda install -y cudatoolkit-dev=11.3 -c conda-forge
pip install tqdm==4.64.0 hydra-core==1.1.2 omegaconf==2.1.2 clean-fid==0.1.23 wandb==0.12.11 ipdb==0.13.9 lpips==0.1.4 einops==0.4.1 inputimeout==1.0.4 pytorch-lightning==1.5.10 matplotlib==3.5.2 "mpl_interactions[jupyter]==0.21.0" protobuf~=3.19.0 moviepy==1.0.3
cd blobgan
python scripts/setup_fid.py

And if you haven't installed ninja yet on your machine (to compile custom C++ code), do that. On Linux, this looks like:

wget -q --show-progress https://github.com/ninja-build/ninja/releases/download/v1.10.2/ninja-linux.zip
sudo unzip -q ninja-linux.zip -d /usr/local/bin/
sudo update-alternatives --install /usr/bin/ninja ninja /usr/local/bin/ninja 1 --force

Running pretrained models

See scripts/load_model.py for an example of how to load a pre-trained model (using the provided load_{blobgan/stylegan}_model functions, which can be called from elsewhere) and generate images with it. You can also run the file from the command line to generate images and save them to disk. For example, from the blobgan directory, you can run:

python scripts/load_model.py --model_data bed --dl_dir models --save_dir out --n_imgs 32 --save_blobs --label_blobs

Or

python scripts/load_model.py --model_name stylegan --model_data conference --truncate 0.4

Note that the first run may take a minute or two longer as custom C++ code is compiled. See the command's help for more details and options: scripts/load_model.py --help

Using these functions, you can access pretrained models on bedrooms (trained with or without jitter); conference rooms; and kitchens, living rooms, and dining rooms.

Training your own model

Before training your model, you'll need to modify src/configs/experiments/local.yaml to include your WandB information and machine-specific configuration (such as path to data -- dataset.path or dataset.basepath -- and number of GPUs trainer.gpus). Alternatively, you can exclude local from the experiments option in the commands below and specify these parameters directly on the command line.

To turn off logging entirely, pass logger=false, or to only log to disk but not write to server, pass wandb.offline=true. Our code currently only supports WandB logging.

Here's an example command which will train a model on LSUN bedrooms. We list the configuration modules to load for this experiment (blobgan, local, jitter) and then specify any other options as we desire. For example, if we wanted to train a model without jitter, we could just remove that module from the experiments array.

python src/run.py +experiment=[blobgan,local,jitter] wandb.name='10-blob BlobGAN on bedrooms' 

In some shells, you may need to add extra quotes around some of these options to prevent them from being parsed immediately on the command line.

Train on the LSUN category of your choice by passing in dataset.category, e.g. dataset.category=church. Tackle multiple categories at once with dataset=multilsun and dataset.categories=[kitchen,bedroom].

You can also train on any collection of images by selecting dataset=imagefolder and passing in the path. The code expects at least a subfolder named train and optional subfolders named validate and test. The below command also illustrates how to set arbitrary options using Hydra syntax, such as turning off FID logging or changing dataloader batch size:

python src/run.py +experiment=[blobgan,local,jitter] wandb.name='20-blob BlobGAN on Places' dataset.dataloader.batch_size=24 +model.log_fid_every_epoch=false dataset=imagefolder +dataset.path=/path/to/places/ model.n_features=20

Other parameters of interest are likely trainer.log_every_n_steps, model.log_images_every_n_steps, and model.log_fid_every_n_steps, which control frequency of logging scalars, images, and FID (set any of the latter two to -1 to disable). Also check out checkpoint.every_n_train_steps and checkpoint.save_top_k which dictate checkpoint saving frequency and decide how many most recent checkpoints to keep (-1 means keep everything).

Changing model feature dimensions

To change d_in, set both model.layout_net.feature_dim and model.generator.override_c_in to the same value. To change d_style, change model.dim.

Logging FID during training and at test

In the initial codebase setup, you should have run scripts/setup_fid.py which will download and install FID statistics for three different datasets:

• Bedrooms: lsun_bedroom
• Conference rooms: lsun_conference
• Kitchens, living rooms, dining rooms: lsun_kld

If either model.log_fid_every_n_steps > -1 or model.log_fid_every_epoch == true, make sure that model.fid_stats_name is passed in. If you are training on one of the three datasets from the paper, just pass in the string from the list above.

If you are training on your own data, you'll need to first run setup_fid.py to precompute statistics on that. The command might look something like:

python scripts/setup_fid.py --action compute_new --path /path/to/new/data --name newdata -j 32 -bs 256

Then, pass model.fid_stats_name=newdata on the command line.

Note that the precomputed FID statistics are on 256px images. You will need to recompute if training at higher resolution.

To run FID logging at test time, a simple snippet such as the following will return the score:

model = load_blobgan_model('bed', 'models', 'cuda', fixed_noise=False)
model.fid_stats_name = 'lsun_bedroom'
model.fid_n_imgs = 50000
print(model.log_fid('train'))

Resuming training

To continue a training run that was terminated, simply add resume.id=PREVIOUS RUN ID. To resume from a previous run but start a new WandB run (e.g. to avoid overwriting previous checkpoints), also pass in wandb.id=null.

Training StyleGAN2

Many of the above command line options apply (for controlling data and logging). For example, to train a StyleGAN2 model on LSUN conference rooms, run:

python src/run.py +experiment=[gan,local] wandb.name='Conference room StyleGAN2' dataset.category=conference

This uses default StyleGAN2 hyperparameters: R1 regularization on D every 16 steps, path length regularization on G every 4, R1 weight 50 or gamma=100 (the weight is gamma/2).

Training inversion encoders

The same is true for training an inversion encoder. See this example command:

python src/run.py +experiment=[invertblobgan,local] wandb.name='Inversion model' +model.G_pretrained.id="BLOBGAN MODEL ID HERE" +model.trunc_min=0.2 +model.trunc_max=0.4 model.lambda.fake_latents_MSE=10

Be sure to specify model.G_pretrained.id to match the ID of the BlobGAN model you are trying to invert. Also, you can set model.G_pretrained.log_dir to tell the program where to look for the model logs (this defaults to ./logs if unspecified). The options trunc_min and trunc_max specify what truncation level to use (randomly sampled within the specified interval) when sampling fake images. If both are set to the same value (including zero, the default), this value will always be used.

Citation

If our code or models aided your research, please cite our paper:

@misc{epstein2022blobgan,
      title={BlobGAN: Spatially Disentangled Scene Representations},
      author={Dave Epstein and Taesung Park and Richard Zhang and Eli Shechtman and Alexei A. Efros},
      year={2022},
      eprint={2205.02837},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}		

Code acknowledgments

This repository is built on top of rosinality's excellent PyTorch re-implementation of StyleGAN2 and Bill Peebles' GANgealing codebase.