Skip to content

[SIGGRAPH'22] StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasets


Notifications You must be signed in to change notification settings


Repository files navigation

This repository contains code for our SIGGRAPH'22 paper "StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasets"

by Axel Sauer, Katja Schwarz, and Andreas Geiger.

If you find our code or paper useful, please cite

  author    = {Axel Sauer and Katja Schwarz and Andreas Geiger},
  title     = {StyleGAN-XL: Scaling StyleGAN to Large Diverse Datasets},
  journal   = {},
  volume    = {abs/2201.00273},
  year      = {2022},
  url       = {},
Rank on Papers With Code  

Related Projects

  • Projected GANs Converge Faster (NeurIPS'21)  -  Official Repo  -  Projected GAN Quickstart
  • StyleGAN-XL + CLIP (Implemented by CasualGANPapers)  -  StyleGAN-XL + CLIP
  • StyleGAN-XL + CLIP (Modified by Katherine Crowson to optimize in W+ space)  -  StyleGAN-XL + CLIP


  • 64-bit Python 3.8 and PyTorch 1.9.0 (or later). See for PyTorch install instructions.
  • CUDA toolkit 11.1 or later.
  • 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 original StyleGAN3 repo and the following issues: #23.
  • Windows user struggling installing the env might find #10 helpful.
  • Use the following commands with Miniconda3 to create and activate your PG Python environment:
    • conda env create -f environment.yml
    • conda activate sgxl

Data Preparation

For a quick start, you can download the few-shot datasets provided by the authors of FastGAN. You can download them here. To prepare the dataset at the respective resolution, run

python --source=./data/pokemon --dest=./data/ \
  --resolution=256x256 --transform=center-crop

You need to follow our progressive growing scheme to get the best results. Therefore, you should prepare separate zips for each training resolution. You can get the datasets we used in our paper at their respective websites (FFHQ, ImageNet).


For progressive growing, we train a stem on low resolution, e.g., 162 pixels. When the stem is finished, i.e., FID is saturating, you can start training the upper stages; we refer to these as superresolution stages.

Training the stem

Training StyleGAN-XL on Pokemon using 8 GPUs:

python --outdir=./training-runs/pokemon --cfg=stylegan3-t --data=./data/ \
    --gpus=8 --batch=64 --mirror=1 --snap 10 --batch-gpu 8 --kimg 10000 --syn_layers 10

--batch specifies the overall batch size, --batch-gpu specifies the batch size per GPU. 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. If you don't want to track metrics, set --metrics=none. You can inspect fid50k_full.json or run tensorboard in training-runs/ to monitor the training progress.

For a class-conditional dataset (ImageNet, CIFAR-10), add the flag --cond True . The dataset needs to contain the class labels; see the StyleGAN2-ADA repo on how to prepare class-conditional datasets.

Training the super-resolution stages

Continuing with pretrained stem:

python --outdir=./training-runs/pokemon --cfg=stylegan3-t --data=./data/ \
  --gpus=8 --batch=64 --mirror=1 --snap 10 --batch-gpu 8 --kimg 10000 --syn_layers 10 \
  --superres --up_factor 2 --head_layers 7 \
  --path_stem training-runs/pokemon/00000-stylegan3-t-pokemon16-gpus8-batch64/best_model.pkl

--up_factor allows to train several stages at once, i.e., with --up_factor=4 and a 162 stem you can directly train at resolution 642.

If you have enough compute, a good tactic is to train several stages in parallel and then restart the superresolution stage training once in a while. The current stage will then reload its previous stem's best_model.pkl. Performance can sometimes drop at first because of domain shift, but the superresolution stage quickly recovers and improves further.

Training recommendations for datasets other than ImageNet

The default settings are tuned for ImageNet. For smaller datasets (<50k images) or well-curated datasets (FFHQ), you can significantly decrease the model size enabling much faster training. Recommended settings are: --cbase 16384 --cmax 256 --syn_layers 7 and for superresolution stages --head_layers 4.

Suppose you want to train as few stages as possible. We recommend training a 32x32 or 64x64 stem, then directly scaling to the final resolution (as described above, you must adjust --up_factor accordingly). However, generally, progressive growing yields better results faster as the throughput is much higher at lower resolutions. This can be seen in this figure by Karras et al., 2017:

Generating Samples & Interpolations

To generate samples and interpolation videos, run

python --outdir=out --trunc=0.7 --seeds=10-15 --batch-sz 1 \


python --output=lerp.mp4 --trunc=0.7 --seeds=0-31 --grid=4x2 \

For class-conditional models, you can pass the class index via --class, a index-to-label dictionary for Imagenet can be found here. For interpolation between classes, provide, e.g., --cls=0-31 to The list of classes has to be the same length as --seeds.

To generate a conditional sample sheet, run

python --outdir=sample_sheets --trunc=1.0 \
  --network= \
  --samples-per-class 4 --classes 0-32 --grid-width 32

For ImageNet models, we enable multi-modal truncation (proposed by Self-Distilled GAN). We generated 600k find 10k cluster centroids via k-means. For a given samples, multi-modal truncation finds the closest centroids and interpolates towards it. To switch from uni-model to multi-modal truncation, pass


No Truncation Uni-Modal Truncation Multi-Modal Truncation

Image Inversion

To invert a given image via latent optimization, and optionally use our reimplementation of Pivotal Tuning Inversion, run

python --outdir=inversion_out \
  --target media/jay.png \
  --inv-steps 1000 --run-pti --pti-steps 350 \

Provide an image via target, it is automatically resized and center-cropped to match the generator network. You do not need to provide a class for ImageNet models, we infer the class of a given sample via a pretrained classifier.

Image Editing

To use our reimplementation of StyleMC, and generate the example above, run

python --outdir=stylemc_out \
  --text-prompt "a chimpanzee | laughter | happyness| happy chimpanzee | happy monkey | smile | grin" \
  --seeds 0-256 --class-idx 367 --layers 10-30 --edit-strength 0.75 --init-seed 49 \
  --network= \

Recommended workflow:

  • Sample images via
  • Pick a sample and use it as the inital image for by providing --init-seed and --class-idx.
  • Find a direction in style space via --text-prompt.
  • Finetune --edit-strength, --layers, and amount of --seeds.
  • Once you found a good setting, provide a larger model via --bigger-network. The script still optimizes the direction for the smaller model, but uses the bigger model for the final output.

Pretrained Models

We provide the following pretrained models (pass the url as PATH_TO_NETWORK_PKL):

Dataset Res FID PATH
ImageNet 162 0.73
ImageNet 322 1.11
ImageNet 642 1.52
ImageNet 1282 1.77
ImageNet 2562 2.26
ImageNet 5122 2.42
ImageNet 10242 2.51
CIFAR10 322 1.85
FFHQ 2562 2.19
FFHQ 5122 2.23
FFHQ 10242 2.02
Pokemon 2562 23.97
Pokemon 5122 23.82
Pokemon 10242 25.47

Quality Metrics

Per default, tracks FID50k during training. To calculate metrics for a specific network snapshot, run

python --metrics=fid50k_full --network=PATH_TO_NETWORK_PKL

To see the available metrics, run

python --help

We provide precomputed FID statistics for all pretrained models:

unzip -d dnnlib/

Further Information

This repo builds on the codebase of StyleGAN3 and our previous project Projected GANs Converge Faster.