>>730
>>730
Fucked around with it for a bit, decapitation_anon is yours and right is mine in pic. Much less fried, lr2.4-e4, telr4e-5, unet2e-4, batch 4, epochs 6, repeats 6. You can look at the training info in additional networks tab to see more details.

Having hundreds of images gives a lot of leeway for training so probably could have baked even longer. Works okay at 1.0 weight, up to 1.2 even but I didn't test it that much. Target tags seem to work at least a little bit if you give them more weight (guillotine:1.3). Probably didn't catch them all since I only had keep_tokens=4.

https://files.catbox.moe/0iggb5.safetensors

Some more notes for self-help:
- enable ss_shuffle_caption and keep_tokens=# (the number of tags from the beginning of the tag file you want to keep per image for every iteration = your targeted activators), shuffling tags makes it less likely to fry/flexible. Same reason you want more epochs than 1 since epochs shuffle the image order around.
- always train on the novelai model, most models and mixes are rooted in it and the lora will work much better with most models this way
- https://rentry.org/RentrySD/#1162-lora-guides-and-info has quite a lot of information, wherever you learned that 100+ repeats is okay is garbage.