VAE and DRAW, Compared

DRAW generates an image the way a person sketches, building it up over several steps with a moving attention window, rather than producing the whole thing in one pass like a standard VAE. This project tested how much that recurrent, attentive process really matters.

I implemented both models from scratch in PyTorch and ran a controlled comparison on MNIST: a vanilla VAE, DRAW without attention, and DRAW with its Gaussian filterbank attention. Everything is seeded and reproducible, with the hardware and library versions documented so the results rebuild exactly.

The work is written up as a short paper in ICLR format, with training curves and generated samples for each configuration.