Type | |
Stats | 1,239 883 |
Reviews | (357) |
Published | Mar 5, 2023 |
Base Model | |
Training | Steps: 5,000 Epochs: 15 |
Trigger Words | custom anon |
This model was trained on 50 images from MrCatJak
He hasn't released his custom model mix, which made me wonder if it's possible to train a LoRA on someone elses generated images. Turns out, it is.
I have his blessing to upload this LoRA, just tag him on twitter if you use this.
Click here for a model/weight comparison grid for V1
Also try these artist tags at different weights. Sometimes they help, sometimes they don't: Franz Xaver Winterhalter, Albert Lynch, Serge Marshennikov
The training set included a lot of his recent images on /sdg/ and a few of his older ones I had saved. I cropped the images and pruned the tags with my homebrew tagset manager (on github).
Most training settings were either eyeballed or left on whatever I had from my previous LoRAs. Base model was AnythingV3.
(some of the example images probably have the wrong lora name, <lora:acnt [custom anon]> is the same as <lora:MrCatJak_style>, tho I'd expect most people to rename it anyway)