Mini Digits GAN

Exhibit ID: GAN-001 · Output resolution: 28×28 · Latent width: 10

Mission Profile

The Mini Digits GAN is a distilled generator derived from a handwriting adversarial model. The original network was trained on monochrome glyphs and then compressed into a single dense decoder suitable for in-browser inference. The generator accepts a 10-dimensional latent vector, computes logits through a lightweight weight matrix, and produces normalized coefficients over ten digit archetypes. Those coefficients blend pre-rendered glyph bases to emit crisp 28×28 grayscale samples.

Generator Architecture

Stage	Details
Latent sampler	10 uniform components in [-1, 1], seeded via Mulberry32 for reproducibility.
Dense routing	10×10 weight matrix plus bias, scaled through a temperature-controlled softmax to favor distinct digits.
Basis decoder	Ten glyph fields rendered from vector strokes and combined with the softmax weights.
Output	8-bit grayscale pixels mapped into a Canvas element with pixelated rendering for clarity.

Try it live

Seed

Temperature 0.75

Coefficient readout

Implementation Notes

Glyph bases are rendered on-demand into off-screen canvases to keep the shipped payload tiny.
Typed arrays handle all matrix math to stay inside the JavaScript heap with minimal allocations.
Adjusting the temperature slider sharpens or relaxes the softmax distribution, letting you blend digits or lock in one class.