Pocket WGAN-GP is a compact class-conditional adversarial image generator for the
8x8 handwritten-digits benchmark. It combines a projection critic, Wasserstein
training with gradient penalty, and an auxiliary class objective. The interactive
Space exposes class, noise seed, and latent temperature.
GANs can report strong-looking averages while collapsing to one prototype. This
project therefore treats class fidelity… See the full description on the dataset page:
https://huggingface.co/datasets/ARotting/pocket-wgan-evaluation.
