ganlive-ffhq

ganlive-ffhq is NVIDIA's 1024×1024 StyleGAN2 trained on faces (FFHQ), converted for ganlive. Play it in your browser with WebGPU, by hand, with MIDI, or from drums. Nothing to install: it downloads once and runs on your graphics card.

ganlive

ganlive explores a GAN's latent space in real time, with dials derived automatically and played by hand, with MIDI, or from drums. It plays FastGAN and StyleGAN2 models.

Use

Open the player in a browser with WebGPU, such as a current Chrome, Edge, Firefox or Safari, and choose ffhq. It plays engine/program.json and engine/weights.bin from this repository, which ganlive convert wrote from ffhq.pt.

To play it in ganlive's desktop app instead, install it as its user guide describes, then:

uv run --no-sync hf download gustavecortal/ganlive-ffhq --local-dir runs/ffhq
uv run --no-sync ganlive play --checkpoint runs/ffhq/engine --console --no-audio --no-midi

Model

• Architecture: StyleGAN2 with a 512-wide latent, output at 1024×1024.
• Weights: NVIDIA's ffhq.pkl from StyleGAN2-ADA, "FFHQ at 1024x1024, trained using original StyleGAN2", imported with ganlive import-stylegan2.
• Speed: on a $350 Intel Arc A770, 40 fps in Edge, and 41 fps in the desktop app.
• Siblings: ganlive-lichen and ganlive-amber, FastGANs trained on my photographs, which run faster.

Training data

FFHQ, 70,000 photographs of faces from Flickr, collected by NVIDIA.

License

The weights are NVIDIA's, under the NVIDIA Source Code License for StyleGAN2-ADA in LICENSE.txt, for non-commercial use only. The FFHQ dataset is under CC BY-NC-SA 4.0.