stylegan-t

[Project] [PDF] [Video]

This repository contains the training code for our paper "StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis". We do not provide pretrained checkpoints.

by Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, Timo Aila

Requirements

• Use the following commands with Miniconda3 to create and activate your environment:

  conda create --name sgt python=3.9
  conda activate sgt
  conda install pytorch=1.9.1 torchvision==0.10.1 pytorch-cuda=11.6 -c pytorch -c nvidia
  pip install -r requirements.txt
  

• GCC 7 or later compilers. The recommended GCC version depends on your CUDA version; see for example, CUDA 11.4 system requirements.
• If you run into problems when setting up the custom CUDA kernels, we refer to the Troubleshooting docs of the StyleGAN3 repo.

Data Preparation

StyleGAN-T can be trained on unconditional and conditional datasets. For small-scale experiments, we recommend zip datasets. When training on datasets with more than 1 million images, we recommend using webdatasets.

Zip Dataset

Zip-Datasets are stored in the same format as in the previous iterations of StyleGAN: uncompressed ZIP archives containing uncompressed PNG files and a metadata file dataset.json for labels. Custom datasets can be created from a folder containing images.

CIFAR-10: Download the CIFAR-10 python version and convert to ZIP archive:

python dataset_tool.py --source downloads/cifar10/cifar-10-python.tar.gz \
    --dest data/cifar10-32x32.zip

FFHQ: Download the Flickr-Faces-HQ dataset as 1024x1024 images and convert to ZIP archive at the same resolution:

python dataset_tool.py --source downloads/ffhq/images1024x1024 \
    --dest data/ffhq1024.zip --resolution 1024x1024

COCO validation set: The COCO validation set is used for tracking zero-shot FID and CLIP score. First, download the COCO meta data. Then, run

python dataset_tool.py --source downloads/captions_val2014.json \
  --dest data/coco_val256.zip --resolution 256x256 --transform center-crop

It is recommend to prepare the dataset zip at the highest possible resolution, e.g. for FFHQ, the zip should contain images with 1024x1024 pixels. When training lower-resolution models, the training script can downsample the images on the fly.

WebDataset

For preparing webdatasets, we used the excellent img2dataset tool. Documentation for downloading different datasets can be found here. For our experiments, we used data from the following sources: CC3M, CC12M, YFFC100m, Redcaps, LAION-aesthetic-6plus.

The joint dataset should have the following structure

joint_dataset/cc3m/0000.tar
joint_dataset/cc3m/0001.tar
...
joint_dataset/laion_6plus/0000.tar
joint_dataset/laion_6plus/0001.tar
...

Training

Training StyleGAN-T with full capacity on MYDATASET.zip at a resolution of 64x64 pixels:

python -m torch.distributed.run --standalone --nproc_per_node 1 train.py \
  --outdir ./training-runs/ --cfg full --data ./data/MYDATASET.zip \
  --img-resolution 64 --batch 128 --batch-gpu 8 --kimg 25000 --metrics fid50k_full

• The above commands can be parallelized across multiple GPUs by adjusting --nproc_per_node.
• --batch specifies the overall batch size, --batch-gpu specifies the batch size per GPU. Be aware that --batch-gpu is also a hyperparameter as the discriminator uses (local) BatchNorm; We generally recommend --batch-gpu of 4 or 8. The training loop will automatically accumulate gradients if you use fewer GPUs until the overall batch size is reached.
• Samples and metrics are saved in outdir. You can inspect METRIC_NAME.json or run tensorboard in training-runs/ to monitor the training progress.
• The generator will be conditional if the dataset contains text labels; otherwise, it will be unconditional.
• For a webdataset comprised of different subsets, the data path should point to the joint parent directory: --data path/to/joint_dataset/

To use the same configuration we used for our ablation study, use --cfg lite. If you want direct control over network parameters, use a custom config. E.g., a smaller models which has 1 residual block, a capacity multiplier of 16384 and a maximum channel count of 256, run

python -m torch.distributed.run --standalone --nproc_per_node 1 train.py \
  --outdir ./training-runs/ --data ./data/MYDATASET.zip \
  --img-resolution 64 --batch 128 --batch-gpu 8 --kimg 25000 --metrics fid50k_full \
  --cfg custom --cbase 16384 --cmax 256 --res-blocks 1

For a description of all input arguments, run python train.py --help

Starting from pretrained checkpoints

If you want to use a previously trained model, you can start from a checkpoint by specifying its path adding --resume PATH_TO_NETWORK_PKL. If you want to continue training from where you left off in a previous run, you can also specify the number of images processed in that run using --resume-kimg XXX, where XXX is that number.

Training modes

By default, all layers of the generator are trained and the CLIP text encoder is frozen. If you want to train only the text encoder, provide --train-mode text-encoder.

If you want to do progressive growing, first train a model at 64x64 pixels. Then provide the path to this pretrained network via -resume, the new target resolution via --img-resolution and use --train-mode freeze64 to freeze the blocks of the 64x64 model and only train the high resolution layers. For example:

python -m torch.distributed.run --standalone --nproc_per_node 1 train.py \
  --outdir ./training-runs/ --data ./data/MYDATASET.zip \
  --img-resolution 512 --batch 128 --batch-gpu 4 --kimg 25000 \
  --cfg lite --resume PATH_TO_NETWORK_64 --train-mode freeze64

Generating Samples

To generate samples with a given network, run

python gen_images.py --network PATH_TO_NETWORK_PKL \
 --prompt 'A painting of a fox in the style of starry night.' \
 --truncation 1 --outdir out --seeds 0-29

For a description of all input arguments, run python gen_images.py --help

Quality Metrics

To calculate metrics for a specific network snapshot, run

python calc_metrics.py --metrics METRIC_NAME --network PATH_TO_NETWORK_PKL

Metric computation is only supported on zip datasets, not webdatasets. The zero-shot COCO metrics expect a coco_val256.zip to be present in the same folder as the training dataset. Alternatively, one can explicitely set an environment variable as follows: export COCOPATH=path/to/coco_val256.zip.

To see the available metrics, run python calc_metrics.py --help

License

Copyright © 2023, NVIDIA Corporation. All rights reserved.

This work is made available under the Nvidia Source Code License.

Excempt are the files training/diffaug.py and networks/vit_utils.py which are partially or fully based on third party github repositories. These two files are copyright their respective authors and under their respective licenses; we include the original license and link to the source at the beginning of the files.

Development

This is a research reference implementation and is treated as a one-time code drop. As such, we do not accept outside code contributions in the form of pull requests.

Citation

@InProceedings{Sauer2023ARXIV,
  author    = {Axel Sauer and Tero Karras and Samuli Laine and Andreas Geiger and Timo Aila},
  title     = {{StyleGAN-T}: Unlocking the Power of {GANs} for Fast Large-Scale Text-to-Image Synthesis},
  journal   = {{arXiv.org}},
  volume    = {abs/2301.09515},
  year      = {2023},
  url       = {https://arxiv.org/abs/2301.09515},
}