Drifting-Models-MNIST

This repository contains a compact MNIST reimplementation of the drifting-model pipeline.

Workflow

1. Train the latent VAE backbone with train_ae.py.
2. Train the latent drift model with train.py.
3. Sample digits with generate.py.

Training

python train_ae.py
python train.py --epochs 30

Generation

python generate.py --digit 7 --strength 1.0 --num_samples 16 --output_path ./outputs/generated_digits.png

Checkpoints are written to ./checkpoints and image grids to ./outputs.

If you trained before this rebuild, retrain both stages. The old generator checkpoint was produced with a collapse-prone objective and will not give reliable digit control.