pocket-wgan

Pocket WGAN-GP

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.

Why this experiment exists

GANs can report strong-looking averages while collapsing to one prototype. This project therefore treats class fidelity, within-class variance, quantized sample uniqueness, nearest-training-example distance, and exact-copy rate as first-class evaluation outputs. A frozen 2,198-parameter classifier trained in Tiny Vision Foundry supplies the label-fidelity measurement.

Verified local result

The 31,712-parameter generator reached 97.0% class fidelity across 1,000 held-out generations. Its mean within-class pixel variance was 98.24% of the real training data baseline. Every quantized sample was unique, the exact training-copy rate was zero, and mean nearest-training-image MSE was 0.03026. The critic contains 17,931 parameters and received 7,200 updates.

Reproduce

uv run python projects/pocket-wgan/train.py
uv run pytest tests/test_pocket_wgan.py

The published bundle contains the generator and critic in SafeTensors format, the complete evaluation report, a 100-sample contact sheet, and 1,000 generated evaluation samples. Reported numbers are local measurements, not claims about a currently deployed Hub Space.