NVIDIA's follow-up GAN fixes artifacts and makes fakes traceable

What it does This is NVIDIA’s official TensorFlow implementation of the StyleGAN2 research paper. It generates high-resolution synthetic images—faces, cars, cats, churches, horses—by re-architecting the original StyleGAN generator to remove characteristic artifacts. The codebase covers training from scratch, loading pretrained networks, projecting photos back into latent space, and computing standard quality metrics.

The interesting bit The authors treated the GAN like a debugging exercise: they identified specific artifacts and capacity bottlenecks, then redesigned generator normalization and revisited progressive growing to fix them. A side effect of the new path-length regularizer is that the generator becomes easier to invert, so you can recover a latent vector from a generated image and reliably detect whether a particular network produced it.

Key highlights

• Pre-trained checkpoints for FFHQ (1024×1024), LSUN Car (512×384), and LSUN Cat, Church, and Horse (256×256)
• Custom TensorFlow operations compiled on-the-fly via NVCC; Linux recommended, Windows supported with Visual Studio in PATH
• Locked to TensorFlow 1.14 or 1.15; TensorFlow 2.x is explicitly unsupported
• Reproducing the full FFHQ paper results at 1024×1024 requires GPUs with at least 16 GB of memory
• Includes projection tools to recover latent vectors from both real and generated images

Caveats

• The dependency stack is frozen circa 2019: TensorFlow 1.x, CUDA 10.0, and cuDNN 7.5
• NVIDIA now points users to StyleGAN2-ADA-PyTorch for new work
• Training times are substantial: roughly seventy days on a single GPU or just under ten days on eight Tesla V100s for the full FFHQ configuration

Verdict Use this if you need the exact TensorFlow implementation for paper reproduction or baseline comparisons. For new projects, NVIDIA’s newer PyTorch release is the pragmatic choice.