High res (+ New model version) (Patreon)

Improved my model further, link here: https://civitai.com/models/94868/stuffy-stuff

------

Overall it’s still pretty close to my last one but with some more improvements like better colors, cleaner generations from txt2img, better at NSFW, and a few more things...

What I changed:

I don’t expect this info to be super useful but I thought it might be interesting.

– I used a better dataset with my latest images.

– Better captioning.

– Tried using regularization images.

– Trained at 1024x1024 resolution.

– Used a different source model.

– I used a lower number of epochs, because last time I think I overtrained it. I also generated sample images during training to better see what was going on at each epoch.

– I made two separate folders for the dataset, one with the best images, which I set to 2 repeats, and another one with the rest, set to one repeat.

– Then I did the same merging steps as my last one. I hope I get to the point where I can reduce the amount of merging and can keep most of the raw trained model instead.

I still need to learn and test a lot of things, I’m also excited to see what can be done with the new SDXL models that will be released in a couple of weeks.