R3GAN2 for WebGPU

The files the R3GAN2 project page loads to generate images live in your browser: the ImageNet 256ร—256 generators (S, M, L) and the FFHQ 256ร—256 generator, repacked for WebGPU, with the decoders that turn their latents into images. They are not meant for PyTorch; the PyTorch checkpoints will be released with the code.

From the NeurIPS 2026 paper Magnitude-preserving Discriminators Enable Efficient GANs, by Nick Yiwen Huang, Jackson Woodleigh*, Aaron Gokaslan*, Xinjie Yi and James Tompkin (*equal contribution).

Files and licences

Files What Licence
imagenet_{s,m,l}_weights.bin, .json, imagenet_{s,m,l}_embed.bin, imagenet_{s,m,l}_test*.bin R3GAN2 ImageNet 256ร—256 generators S (17 M parameters), M (61 M) and L (131 M), class embeddings, and reference outputs for the demo's self-check CC BY-NC 4.0; the TAEF2 decoder weights packed into the .bin files are MIT (see below)
ffhq_weights.bin, .json, ffhq_test.bin R3GAN2 FFHQ 256ร—256 generator (pixel space) CC BY-NC-SA 4.0, following FFHQ's licence
vae_dec.onnx The FLUX.2 autoencoder's decoder (Black Forest Labs), exported to ONNX Apache 2.0; our changes are listed in NOTICE
(inside imagenet_*_weights.bin) The TAEF2 decoder by Ollin Boer Bohan, repacked for WebGPU MIT

The ImageNet generators work in the latent space of the FLUX.2 autoencoder; the demo decodes their latents with TAEF2 (fast) or the full FLUX.2 decoder.

Citation

@inproceedings{huang2026magnitude,
  title     = {Magnitude-preserving Discriminators Enable Efficient {GAN}s},
  author    = {Huang, Nick Yiwen and Woodleigh, Jackson and Gokaslan, Aaron and Yi, Xinjie and Tompkin, James},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support