I built an MNIST for video diffusion — and trained a 5-second dancer on a free Colab

Every one of these dancers was drawn from pure noise by a model that trained for a few hours on one GPU. And for every frame in the training data, we know exactly where each arm and leg really is — so