SOTA GANs are hard to train and to explore, and StyleGAN2/ADA/3 are no different. The point of this repository is to allow the user to both easily train and explore the trained models without unnecessary headaches.
As before, we will build upon the official repository, which has the advantage of being backwards-compatible. As such, we can use our previously-trained models from StyleGAN2 and StyleGAN2-ADA. Please get acquainted with the official repository and its codebase, as we will be building upon it and as such, increase its capabilities (but hopefully not its complexity!).
This repository adds the following (not yet the complete list):
- Dataset tool
- Add
--center-crop-tall
: add vertical black bars to the sides instead, in the same vein as the horizontal bars in--center-crop-wide
. - Grayscale images in the dataset are converted to
RGB
. - If the dataset tool encounters an error, print it along the offending image, but continue with the rest of the dataset (pull #39 from Andreas Jansson).
- TODO: Add multi-crop, as used in Earth View.
- Add
- Training
--mirrory
: Added vertical mirroring for doubling the dataset size--gamma
: If no R1 regularization is provided, the heuristic formula will be used from StyleGAN2.--augpipe
: Now available to use is StyleGAN2-ADA's full list of augpipe, e,g.,blit
,geom
,bgc
,bgcfnc
, etc.--img-snap
: When to save snapshot images, so now it's independent of when the model is saved;--snap-res
: The resolution of the snapshots, depending on your screen resolution, or how many images you wish to see per tick. Available resolutions:1080p
,4k
, and8k
.--resume-kimg
: Starting number ofkimg
, useful when continuing training a previous run--outdir
: Automatically set astraining-runs
--metrics
: Now set by default toNone
, so there's no need to worry about this one--resume
: All available pre-trained models from NVIDIA can be found with a simple dictionary, depending on the--cfg
used. For example, if--cfg=stylegan3-r
, then to transfer learn from FFHQU at 1024 resolution, set--resume=ffhqu1024
. TODO: finish the following table, but full list available here.
Available Models | ffhq256 |
ffhqu256 |
ffhq512 |
ffhq1024 |
ffhqu1024 |
stylegan2 |
✔️ | ✔️ | ✔️ | ✔️ | ✔️ |
stylegan3-t |
✔️ | ✔️ | ✔️ | ||
stylegan3-r |
✔️ | ✔️ | ✔️ |
- Interpolation videos
- Random interpolation
- Style-mixing
- Sightseeding
- Circular interpolation
- Visual-reactive interpolation (Beta)
- Projection into the latent space
- Project into W+
- Additional losses to use for better projection (e.g., using VGG16 or CLIP)
- Discriminator Synthesis (official code)
- Generate a static image or a video with a feedback loop
- Start from a random image (
random
orperlin
, using Mathieu Duchesneau's implementation) or from an existing one
- Expansion on GUI/
visualizer.py
- Added the rest of the affine transformations
- General model and code additions
- No longer necessary to specify
--outdir
when running the code, as the output directory will be automatically generated - Better sampling? (TODO)
- StyleGAN3: anchor the latent space for easier to follow interpolations
- No longer necessary to specify
TODO: Finish documentation for better user experience, add videos/images, code samples.
Alias-Free Generative Adversarial Networks (StyleGAN3)
Official PyTorch implementation of the NeurIPS 2021 paper
Alias-Free Generative Adversarial Networks
Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, Timo Aila
https://nvlabs.github.io/stylegan3
Abstract: We observe that despite their hierarchical convolutional nature, the synthesis process of typical generative adversarial networks depends on absolute pixel coordinates in an unhealthy manner. This manifests itself as, e.g., detail appearing to be glued to image coordinates instead of the surfaces of depicted objects. We trace the root cause to careless signal processing that causes aliasing in the generator network. Interpreting all signals in the network as continuous, we derive generally applicable, small architectural changes that guarantee that unwanted information cannot leak into the hierarchical synthesis process. The resulting networks match the FID of StyleGAN2 but differ dramatically in their internal representations, and they are fully equivariant to translation and rotation even at subpixel scales. Our results pave the way for generative models better suited for video and animation.
For business inquiries, please visit our website and submit the form: NVIDIA Research Licensing
This repository is an updated version of stylegan2-ada-pytorch, with several new features:
- Alias-free generator architecture and training configurations (
stylegan3-t
,stylegan3-r
). - Tools for interactive visualization (
visualizer.py
), spectral analysis (avg_spectra.py
), and video generation (gen_video.py
). - Equivariance metrics (
eqt50k_int
,eqt50k_frac
,eqr50k
). - General improvements: reduced memory usage, slightly faster training, bug fixes.
Compatibility:
- Compatible with old network pickles created using stylegan2-ada and stylegan2-ada-pytorch. (Note: running old StyleGAN2 models on StyleGAN3 code will produce the same results as running them on stylegan2-ada/stylegan2-ada-pytorch. To benefit from the StyleGAN3 architecture, you need to retrain.)
- Supports old StyleGAN2 training configurations, including ADA and transfer learning. See Training configurations for details.
- Improved compatibility with Ampere GPUs and newer versions of PyTorch, CuDNN, etc.
While new generator approaches enable new media synthesis capabilities, they may also present a new challenge for AI forensics algorithms for detection and attribution of synthetic media. In collaboration with digital forensic researchers participating in DARPA's SemaFor program, we curated a synthetic image dataset that allowed the researchers to test and validate the performance of their image detectors in advance of the public release. Please see here for more details.
- Result videos
- Curated example images
- StyleGAN3 pre-trained models for config T (translation equiv.) and config R (translation and rotation equiv.)
Access individual networks via
https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/<MODEL>
, where<MODEL>
is one of:
stylegan3-t-ffhq-1024x1024.pkl
,stylegan3-t-ffhqu-1024x1024.pkl
,stylegan3-t-ffhqu-256x256.pkl
stylegan3-r-ffhq-1024x1024.pkl
,stylegan3-r-ffhqu-1024x1024.pkl
,stylegan3-r-ffhqu-256x256.pkl
stylegan3-t-metfaces-1024x1024.pkl
,stylegan3-t-metfacesu-1024x1024.pkl
stylegan3-r-metfaces-1024x1024.pkl
,stylegan3-r-metfacesu-1024x1024.pkl
stylegan3-t-afhqv2-512x512.pkl
stylegan3-r-afhqv2-512x512.pkl
- StyleGAN2 pre-trained models compatible with this codebase
Access individual networks via
https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan2/versions/1/files/<MODEL>
, where<MODEL>
is one of:
stylegan2-ffhq-1024x1024.pkl
,stylegan2-ffhq-512x512.pkl
,stylegan2-ffhq-256x256.pkl
stylegan2-ffhqu-1024x1024.pkl
,stylegan2-ffhqu-256x256.pkl
stylegan2-metfaces-1024x1024.pkl
,stylegan2-metfacesu-1024x1024.pkl
stylegan2-afhqv2-512x512.pkl
stylegan2-afhqcat-512x512.pkl
,stylegan2-afhqdog-512x512.pkl
,stylegan2-afhqwild-512x512.pkl
stylegan2-brecahad-512x512.pkl
,stylegan2-cifar10-32x32.pkl
stylegan2-celebahq-256x256.pkl
,stylegan2-lsundog-256x256.pkl
- Linux and Windows are supported, but we recommend Linux for performance and compatibility reasons.
- 1–8 high-end NVIDIA GPUs with at least 12 GB of memory. We have done all testing and development using Tesla V100 and A100 GPUs.
- 64-bit Python 3.8 and PyTorch 1.9.0 (or later). See https://pytorch.org for PyTorch install instructions.
- CUDA toolkit 11.1 or later. (Why is a separate CUDA toolkit installation required? See Troubleshooting).
- GCC 7 or later (Linux) or Visual Studio (Windows) compilers. Recommended GCC version depends on CUDA version, see for example CUDA 11.4 system requirements.
- Python libraries: see environment.yml for exact library dependencies. You can use the following commands with Miniconda3 to create and activate your StyleGAN3 Python environment:
conda env create -f environment.yml
conda activate stylegan3
- Docker users:
- Ensure you have correctly installed the NVIDIA container runtime.
- Use the provided Dockerfile to build an image with the required library dependencies.
The code relies heavily on custom PyTorch extensions that are compiled on the fly using NVCC. On Windows, the compilation requires Microsoft Visual Studio. We recommend installing Visual Studio Community Edition and adding it into PATH
using "C:\Program Files (x86)\Microsoft Visual Studio\<VERSION>\Community\VC\Auxiliary\Build\vcvars64.bat"
.
See Troubleshooting for help on common installation and run-time problems.
Pre-trained networks are stored as *.pkl
files that can be referenced using local filenames or URLs:
# Generate an image using pre-trained AFHQv2 model ("Ours" in Figure 1, left).
python gen_images.py --outdir=out --trunc=1 --seeds=2 \
--network=https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/stylegan3-r-afhqv2-512x512.pkl
# Render a 4x2 grid of interpolations for seeds 0 through 31.
python gen_video.py --output=lerp.mp4 --trunc=1 --seeds=0-31 --grid=4x2 \
--network=https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/stylegan3-r-afhqv2-512x512.pkl
Outputs from the above commands are placed under out/*.png
, controlled by --outdir
. Downloaded network pickles are cached under $HOME/.cache/dnnlib
, which can be overridden by setting the DNNLIB_CACHE_DIR
environment variable. The default PyTorch extension build directory is $HOME/.cache/torch_extensions
, which can be overridden by setting TORCH_EXTENSIONS_DIR
.
Docker: You can run the above curated image example using Docker as follows:
# Build the stylegan3:latest image
docker build --tag stylegan3 .
# Run the gen_images.py script using Docker:
docker run --gpus all -it --rm --user $(id -u):$(id -g) \
-v `pwd`:/scratch --workdir /scratch -e HOME=/scratch \
stylegan3 \
python gen_images.py --outdir=out --trunc=1 --seeds=2 \
--network=https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/stylegan3-r-afhqv2-512x512.pkl
Note: The Docker image requires NVIDIA driver release r470
or later.
The docker run
invocation may look daunting, so let's unpack its contents here:
--gpus all -it --rm --user $(id -u):$(id -g)
: with all GPUs enabled, run an interactive session with current user's UID/GID to avoid Docker writing files as root.-v `pwd`:/scratch --workdir /scratch
: mount current running dir (e.g., the top of this git repo on your host machine) to/scratch
in the container and use that as the current working dir.-e HOME=/scratch
: let PyTorch and StyleGAN3 code know where to cache temporary files such as pre-trained models and custom PyTorch extension build results. Note: if you want more fine-grained control, you can instead setTORCH_EXTENSIONS_DIR
(for custom extensions build dir) andDNNLIB_CACHE_DIR
(for pre-trained model download cache). You want these cache dirs to reside on persistent volumes so that their contents are retained across multipledocker run
invocations.
This release contains an interactive model visualization tool that can be used to explore various characteristics of a trained model. To start it, run:
python visualizer.py
You can use pre-trained networks in your own Python code as follows:
with open('ffhq.pkl', 'rb') as f:
G = pickle.load(f)['G_ema'].cuda() # torch.nn.Module
z = torch.randn([1, G.z_dim]).cuda() # latent codes
c = None # class labels (not used in this example)
img = G(z, c) # NCHW, float32, dynamic range [-1, +1], no truncation
The above code requires torch_utils
and dnnlib
to be accessible via PYTHONPATH
. It does not need source code for the networks themselves — their class definitions are loaded from the pickle via torch_utils.persistence
.
The pickle contains three networks. 'G'
and 'D'
are instantaneous snapshots taken during training, and 'G_ema'
represents a moving average of the generator weights over several training steps. The networks are regular instances of torch.nn.Module
, with all of their parameters and buffers placed on the CPU at import and gradient computation disabled by default.
The generator consists of two submodules, G.mapping
and G.synthesis
, that can be executed separately. They also support various additional options:
w = G.mapping(z, c, truncation_psi=0.5, truncation_cutoff=8)
img = G.synthesis(w, noise_mode='const', force_fp32=True)
Please refer to gen_images.py
for complete code example.
Datasets are stored as 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; see python dataset_tool.py --help
for more information. Alternatively, the folder can also be used directly as a dataset, without running it through dataset_tool.py
first, but doing so may lead to suboptimal performance.
FFHQ: Download the Flickr-Faces-HQ dataset as 1024x1024 images and create a zip archive using dataset_tool.py
:
# Original 1024x1024 resolution.
python dataset_tool.py --source=/tmp/images1024x1024 --dest=~/datasets/ffhq-1024x1024.zip
# Scaled down 256x256 resolution.
python dataset_tool.py --source=/tmp/images1024x1024 --dest=~/datasets/ffhq-256x256.zip \
--resolution=256x256
See the FFHQ README for information on how to obtain the unaligned FFHQ dataset images. Use the same steps as above to create a ZIP archive for training and validation.
MetFaces: Download the MetFaces dataset and create a ZIP archive:
python dataset_tool.py --source=~/downloads/metfaces/images --dest=~/datasets/metfaces-1024x1024.zip
See the MetFaces README for information on how to obtain the unaligned MetFaces dataset images. Use the same steps as above to create a ZIP archive for training and validation.
AFHQv2: Download the AFHQv2 dataset and create a ZIP archive:
python dataset_tool.py --source=~/downloads/afhqv2 --dest=~/datasets/afhqv2-512x512.zip
Note that the above command creates a single combined dataset using all images of all three classes (cats, dogs, and wild animals), matching the setup used in the StyleGAN3 paper. Alternatively, you can also create a separate dataset for each class:
python dataset_tool.py --source=~/downloads/afhqv2/train/cat --dest=~/datasets/afhqv2cat-512x512.zip
python dataset_tool.py --source=~/downloads/afhqv2/train/dog --dest=~/datasets/afhqv2dog-512x512.zip
python dataset_tool.py --source=~/downloads/afhqv2/train/wild --dest=~/datasets/afhqv2wild-512x512.zip
You can train new networks using train.py
. For example:
# Train StyleGAN3-T for AFHQv2 using 8 GPUs.
python train.py --outdir=~/training-runs --cfg=stylegan3-t --data=~/datasets/afhqv2-512x512.zip \
--gpus=8 --batch=32 --gamma=8.2 --mirror=1
# Fine-tune StyleGAN3-R for MetFaces-U using 1 GPU, starting from the pre-trained FFHQ-U pickle.
python train.py --outdir=~/training-runs --cfg=stylegan3-r --data=~/datasets/metfacesu-1024x1024.zip \
--gpus=8 --batch=32 --gamma=6.6 --mirror=1 --kimg=5000 --snap=5 \
--resume=https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/stylegan3-r-ffhqu-1024x1024.pkl
# Train StyleGAN2 for FFHQ at 1024x1024 resolution using 8 GPUs.
python train.py --outdir=~/training-runs --cfg=stylegan2 --data=~/datasets/ffhq-1024x1024.zip \
--gpus=8 --batch=32 --gamma=10 --mirror=1 --aug=noaug
Note that the result quality and training time depend heavily on the exact set of options. The most important ones (--gpus
, --batch
, and --gamma
) must be specified explicitly, and they should be selected with care. See python train.py --help
for the full list of options and Training configurations for general guidelines & recommendations, along with the expected training speed & memory usage in different scenarios.
The results of each training run are saved to a newly created directory, for example ~/training-runs/00000-stylegan3-t-afhqv2-512x512-gpus8-batch32-gamma8.2
. The training loop exports network pickles (network-snapshot-<KIMG>.pkl
) and random image grids (fakes<KIMG>.png
) at regular intervals (controlled by --snap
). For each exported pickle, it evaluates FID (controlled by --metrics
) and logs the result in metric-fid50k_full.jsonl
. It also records various statistics in training_stats.jsonl
, as well as *.tfevents
if TensorBoard is installed.
By default, train.py
automatically computes FID for each network pickle exported during training. We recommend inspecting metric-fid50k_full.jsonl
(or TensorBoard) at regular intervals to monitor the training progress. When desired, the automatic computation can be disabled with --metrics=none
to speed up the training slightly.
Additional quality metrics can also be computed after the training:
# Previous training run: look up options automatically, save result to JSONL file.
python calc_metrics.py --metrics=eqt50k_int,eqr50k \
--network=~/training-runs/00000-stylegan3-r-mydataset/network-snapshot-000000.pkl
# Pre-trained network pickle: specify dataset explicitly, print result to stdout.
python calc_metrics.py --metrics=fid50k_full --data=~/datasets/ffhq-1024x1024.zip --mirror=1 \
--network=https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/stylegan3-t-ffhq-1024x1024.pkl
The first example looks up the training configuration and performs the same operation as if --metrics=eqt50k_int,eqr50k
had been specified during training. The second example downloads a pre-trained network pickle, in which case the values of --data
and --mirror
must be specified explicitly.
Note that the metrics can be quite expensive to compute (up to 1h), and many of them have an additional one-off cost for each new dataset (up to 30min). Also note that the evaluation is done using a different random seed each time, so the results will vary if the same metric is computed multiple times.
Recommended metrics:
fid50k_full
: Fréchet inception distance[1] against the full dataset.kid50k_full
: Kernel inception distance[2] against the full dataset.pr50k3_full
: Precision and recall[3] againt the full dataset.ppl2_wend
: Perceptual path length[4] in W, endpoints, full image.eqt50k_int
: Equivariance[5] w.r.t. integer translation (EQ-T).eqt50k_frac
: Equivariance w.r.t. fractional translation (EQ-Tfrac).eqr50k
: Equivariance w.r.t. rotation (EQ-R).
Legacy metrics:
fid50k
: Fréchet inception distance against 50k real images.kid50k
: Kernel inception distance against 50k real images.pr50k3
: Precision and recall against 50k real images.is50k
: Inception score[6] for CIFAR-10.
References:
- GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium, Heusel et al. 2017
- Demystifying MMD GANs, Bińkowski et al. 2018
- Improved Precision and Recall Metric for Assessing Generative Models, Kynkäänniemi et al. 2019
- A Style-Based Generator Architecture for Generative Adversarial Networks, Karras et al. 2018
- Alias-Free Generative Adversarial Networks, Karras et al. 2021
- Improved Techniques for Training GANs, Salimans et al. 2016
The easiest way to inspect the spectral properties of a given generator is to use the built-in FFT mode in visualizer.py
. In addition, you can visualize average 2D power spectra (Appendix A, Figure 15) as follows:
# Calculate dataset mean and std, needed in subsequent steps.
python avg_spectra.py stats --source=~/datasets/ffhq-1024x1024.zip
# Calculate average spectrum for the training data.
python avg_spectra.py calc --source=~/datasets/ffhq-1024x1024.zip \
--dest=tmp/training-data.npz --mean=112.684 --std=69.509
# Calculate average spectrum for a pre-trained generator.
python avg_spectra.py calc \
--source=https://api.ngc.nvidia.com/v2/models/nvidia/research/stylegan3/versions/1/files/stylegan3-r-ffhq-1024x1024.pkl \
--dest=tmp/stylegan3-r.npz --mean=112.684 --std=69.509 --num=70000
# Display results.
python avg_spectra.py heatmap tmp/training-data.npz
python avg_spectra.py heatmap tmp/stylegan3-r.npz
python avg_spectra.py slices tmp/training-data.npz tmp/stylegan3-r.npz
Copyright © 2021, NVIDIA Corporation & affiliates. All rights reserved.
This work is made available under the Nvidia Source Code License.
@inproceedings{Karras2021,
author = {Tero Karras and Miika Aittala and Samuli Laine and Erik H\"ark\"onen and Janne Hellsten and Jaakko Lehtinen and Timo Aila},
title = {Alias-Free Generative Adversarial Networks},
booktitle = {Proc. NeurIPS},
year = {2021}
}
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.
We thank David Luebke, Ming-Yu Liu, Koki Nagano, Tuomas Kynkäänniemi, and Timo Viitanen for reviewing early drafts and helpful suggestions. Frédo Durand for early discussions. Tero Kuosmanen for maintaining our compute infrastructure. AFHQ authors for an updated version of their dataset. Getty Images for the training images in the Beaches dataset. We did not receive external funding or additional revenues for this project.