NotSHAFT

Retrained the model. This time, there is some special sauce (Which I call crop augmentation). Essentially, I ran the images through a model to get the saliency for each model and cropped them to a 1:1 ratio. In my initIal testing, square pictures are now much more defined and close up and the over quality appears to have improved. I described a bit of my process on twitter. The epoch uploaded is epoch 4, which isn't the final epoch but is so far the best performing one. N.B. the trigger word "notshaft style", doesn't strongly affect the image too much, I messed up and forgot to add keep tokens. N.B 2: Consider letting the model stew a bit longer to get more the the fine details out (More steps. Euler a with 80 seems like a decent spot). Also, set the text encoder to 0.5. I'll probably find out how to properly train a text encoder eventually lol. Hint: The examples are your friend. :3