Reddit user builds tiny Stable Diffusion clone from scratch

View on Reddit

The model generates 48x48 images using a bidirectional transformer encoder, trained on Flickr8k and ImageNet. Current loss is 1.1443, early in training.