generators-with-stylegan2: StyleGAN2 face generators you download and run yourself

What generators-with-stylegan2 actually ships

This is not a training framework. It is a set of pretrained StyleGAN2 networks wrapped in the original NVIDIA code layout, plus a small editor for moving faces along learned attribute directions. The README lists five open generators (wanghong, star, model, babies and yellow) rebuilt on StyleGAN2, and two more, mixed-blood and Asian beauty, that it explicitly labels non-open-source. The mixed-blood one is described as bought out and no longer for sale. That split matters more than any feature list: what you can actually download is five faces, not seven.

The intended audience is stated plainly. The README says the project is meant to help people working in film and television, advertising, games and medical aesthetics, and hobbyists. The pitch is cost: generated faces substitute for finding real people to photograph. Every face is synthetic, and the README stresses that each one is a person who does not exist.

The repository layout backs this up. main.py and pretrained_networks.py sit at the top level, with dnnlib/, networks/ and latent_directions/ as the supporting machinery, and edit_photo.py as the attribute editor. There is a Windows installation guide shipped as a .docx file rather than as Markdown, which tells you something about where the author expected friction to land.

How the StyleGAN2 generator pipeline is wired here

The architecture is NVIDIA's StyleGAN2, unchanged in its essentials. A mapping network turns a latent vector into a style vector, which is injected at multiple resolutions of the synthesis network, so coarse layers control pose and face shape while finer layers control texture and colour. The README's stated benefit of the StyleGAN2 base over the earlier version is the removal of the droplet artifacts and the distorted or broken regions that plagued StyleGAN1 output, which it says pushes generation success close to 100 percent and makes bulk generation practical.

Generation is driven from main.py, which takes a network pickle and a seed and writes images. The network pickles are resolved by pretrained_networks.py, which is where the download URLs for each face model live. This is the file to read before anything else, because it determines whether you get a working generator or a 404.

The attribute editor is a separate path. latent_directions/ holds direction vectors, and edit_photo.py applies them to an existing image by projecting it into the latent space and then moving it along a chosen direction. That is a different operation from random sampling: you need a source image, and the quality of the edit depends on how well the projection lands.

Installing it and generating a first batch of faces

The environment requirements are strict and dated, and the README states them without hedging. Python 3.6, 64-bit, with numpy 1.14.3 or newer; TensorFlow 1.14 or 1.15 with GPU support; CUDA 10.0 and cuDNN 7.5; and one or more high-end NVIDIA GPUs. The README is explicit that the code does not support TensorFlow 2.0, and that on Windows you must use TensorFlow 1.14 because 1.15 will not work. The author's own tested configuration is listed as Windows 10, a GTX 1050 Ti, CUDA 10.0, cuDNN 7.6.5, tensorflow-gpu 1.14.0 and Visual Studio 2017.

The README points Docker users at the Dockerfile from the NVlabs/stylegan2 repository rather than shipping one here, so containerised setup means borrowing NVIDIA's build.

On Windows, the compilation step needs Microsoft Visual Studio on PATH. The README recommends Visual Studio Community Edition and adding it with the vcvars64.bat script, quoting this path:

C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Auxiliary\Build\vcvars64.bat

If that is missing you get the error the README devotes a whole section to: Could not find MSVC/GCC/CLANG installation on this computer. The documented fix is to make sure the C++ desktop development workload is selected during Visual Studio installation.

With the environment in place, generation runs through main.py against a pretrained network. The README does not print a command line for it, so open main.py and read its argument parser for the flags your checkout expects before running anything. The README states that a purely random, unfiltered set of ten thousand images per generator is published on Baidu Pan links, with extraction codes listed next to each one, which is the fastest way to judge output quality without setting up the environment at all.

Where this project breaks, and who should walk away

The dependency stack is the first wall. TensorFlow 1.14 and 1.15 are years behind current releases, CUDA 10.0 pins you to older NVIDIA drivers, and Python 3.6 is past end of life. If your organisation has standardised on TensorFlow 2.x, this repository is not a candidate without a port, and the README states outright that the code does not support TensorFlow 2.0.

The second wall is memory. The README says that reproducing the paper's results needs an NVIDIA GPU with at least 16 GB of DRAM. The author's own 1050 Ti is a 4 GB card, which the README presents as a working test environment, so the 16 GB figure is about matching published results rather than about running the code at all. Those are different claims, and the README does not separate them. If you plan to generate at high resolution in volume, budget for the larger card.

The third issue is licensing. The repository's licence is reported as NOASSERTION, meaning GitHub could not classify the LICENSE.txt file. The README says the models are copyright www.seeprettyface.com and are fully open for use within reasonable limits, which is not a licence text. Commercial use of the output is exactly the case where you would want an actual grant, and there is no clear one here.

Finally, two of the advertised generators are not available. If your reason for arriving is the mixed-blood or Asian beauty model, the README closes that door itself.

StyleGAN2-ADA and StyleGAN2-pytorch as the real alternatives

The honest comparison is not against another face generator but against the codebases that let you train your own. StyleGAN2-ADA, from NVIDIA, adds adaptive discriminator augmentation, which is designed to make training work with small datasets where the original StyleGAN2 would overfit. That is a different proposition entirely: this repository gives you five finished faces, while StyleGAN2-ADA gives you the ability to produce a face model for a category nobody has trained yet, at the cost of a training run and a dataset.

StyleGAN2-pytorch is the other route. It reimplements the architecture in PyTorch, which removes the TensorFlow 1.x and CUDA 10.0 constraint that dominates setup here, and it has its own ecosystem of pretrained checkpoints. If your team already works in PyTorch, the migration cost of adopting this repository is likely higher than the cost of using a PyTorch implementation with a comparable checkpoint.

The trade-off is straightforward. This project wins when you want images today and one of its five face categories fits. It loses the moment you need a category it does not cover, because it ships no training script and no dataset, only the inference side.

Maintenance, upgrade cost and what the licence does not say

The last push to the repository was on 2026-09-03, and the repository is not archived. The README itself announces that the author will stop making new face generators and move on to other generation work, naming Video-Auto-Wipe as the next direction. That is a statement about scope, not about the code, but it sets expectations: the five open models are the finished set.

Upgrade cost is dominated by the TensorFlow pin. Moving this code to a current TensorFlow would mean replacing the custom ops in the networks/ and dnnlib/ directories, which are compiled against CUDA 10.0. There is no sign in the README of a plan to do that. In practice you freeze the environment, containerise it, and treat it as a fixed artifact.

The licence situation is genuinely unresolved. NOASSERTION means the LICENSE.txt file did not match a known template. The README's own wording, that the models are copyright www.seeprettyface.com and open for use in a reasonable range, is a permission statement without terms. If you intend to publish generated faces or use them in advertising, that ambiguity is the thing to resolve before the images leave your machine, and it is a question for a lawyer rather than for this article.

Editorial conclusion

Adopt it if you need ready-made portrait material and can put a CUDA 10.0, TensorFlow 1.14 or 1.15 machine with a 16 GB NVIDIA GPU in front of it; the model files, not the code, are the product. Do not adopt it if you are on TensorFlow 2.x, on a laptop GPU, or if you need the mixed-blood or Asian beauty generators, which the README describes as not open source and, in the mixed-blood case, bought out. Verify first that the pretrained network files download from the URLs in pretrained_networks.py and that your GPU memory actually fits the resolution you intend, because the README states 16 GB DRAM is the figure for reproducing the paper's results.

Frequently asked questions

Official sources

1. a312863063/generators-with-stylegan2 on GitHub
2. Issues
3. README