USE THE NEW EMBEDDING!!!
Completed.
Although the model is far from perfect, this is where this particular experiment ends.
I'll start over with a different training method.
The final version can handle decently 1024x1024 resolution for most of its concepts.
Right now I don't feel like posting the "keywords" the model was trained on, maybe I'll suffice later.
Description
Less underfit, remember to use the embeddings and if you are confused as to why it's 512 when it's listed 768, it's because I took a base 768 and fine-tuned on 512 ratio for speed.
FAQ
Comments (8)
On my end it works nicely in colab, until I upload the embeddings I recommend using as negative prompt "missing negative prompt", I don't know if it can work for you.
Don't mention the name of the c*lab please, we want to keep it this time
Sorry, I'll leave it off
Fixed.
I'll upload a less underfit model version in a second.
MAN.
If you find interest in this model, please provide sample images because I'm using a technique of fine-tuning that uses the generated images as part of UNET training. Any actually good image can be used for improving the model.
For your INFO: the model has a decent understanding of celebrities, the text encoder is fully uncensored but i did some mess on it's idea of Debbie Harry, Kate Pierson and some more, I'll fix it sooner or later. The model was also trained on some videogame character, just as you can see from the sample pictures, they were not present in the text encoder and I forced them in, unfortunately is a little instable on these characters (turns out that for training the text encoder on characters, it's better to use one single tag)