JakeDiffusion

These images were created using a heavily modified version of the Fast Stable Diffusion notebook by TheLastBen in 2022: GitHub Link. Contact me if you want the weights, but modern solutions probably work better out of the box than anything that I can produce myself on consumer grade hardware.

"Example output"

The model was trained on (512 × 512) selfies of myself, producing what I call "DeepJakes".

"Example output"

Note: Does not do eyes or hands very well. This is probably a result of problems inherent to the version of stable diffusion that I was using and the poor quality of the training data utilised. Does not process expressions other than mild indifference very well because they were not included in the training set. Asking for a portait of a smiling [TOKEN_NAME] produced horrifying results.

Scary!