Mycsina/mnist-conditional-gan

Conditional MNIST GAN

Label-conditioned generator for 28x28 MNIST digits.

This repository contains generator weights only, as required by the assignment. The model uses a 100-dimensional standard-normal latent vector. See model.py for the exact architecture and training-metrics.json for the complete loss history.

Training

• Dataset: ylecun/mnist
• Epochs: 25
• Batch size: 256
• Seed: 42
• Fixed-sample pixel standard deviation: 0.6101

Conditional evaluation

• Independent classifier accuracy on real MNIST test data: 98.51%
• Requested-label agreement on 2000 generated samples: 99.55%

Weights are stored in safetensors format. The model generates synthetic images and can produce malformed or ambiguous samples; it is intended for coursework and experimentation.