StyleGAN3
Image GenerationNVIDIA · StyleGAN · v3.0 · released
About
StyleGAN3 is NVIDIA Labs' third-generation generative adversarial network for photorealistic image synthesis. Introduced alias-free generator design that eliminated the rotation/translation artifacts present in prior StyleGAN versions. Research-grade — non-commercial license.
Intended use: Research on generative image models, style transfer, latent-space manipulation. Pre-trained checkpoints for FFHQ-U (human faces), AFHQv2 (animal faces), MetFaces-U (art portraits). Inference on a single consumer GPU.
Architecture
- Type
- gan
- Parameters
- 25M
Unconditional generative adversarial network with alias-free transformation layers. Generator is a fully-equivariant CNN that produces 1024×1024 RGB images from a 512-dim Gaussian latent code; discriminator is a CNN with progressive growing. Solved the texture-sticking and aliasing artifacts that plagued StyleGAN2 — outputs are translation- and rotation-equivariant under animation. Trained on FFHQ-U, AFHQv2, MetFaces-U corpora.
Memory
- Weights (BF16)
- 0.05 GB
- Activation estimate
- 0.40 GB
Pricing
Free — open weights
Self-host on your own GPU. The calculator surfaces GPU-hours cost on the hardware page instead of an API price.
Provenance
- Source
- nvlabs.github.io
- License
- nvidia-source-code-license-nc
- Hugging Face
- stabilityai/stylegan3
- Last verified
- 2026-06-25