NO restriction for this LoRA! Feel Free to use! (include commercial use for design or something)
The script for making dataset is available in Github.
https://github.com/paseri3739/fuXing_scraper
For more training parameter details, check the binary in hex editor.
Version:
v0.3.1
full_bf16 training version. file size is half of 0.3. In most case, The output is same as v0.3.
v0.4.0
Add more dataset. no full_bf16, 17 epochs because it's not overfitted (especially hands). Maybe eye detail is improved?
v0.4.1
Learning resolution 512 -> 768 to improve the detail of fur.
AdamW8bit -> AdamW to improve responsiveness and accuracy to prompts.
batch 6 -> 2
Recommended positive prompt:
fursuit, furry, realistic, no humans, animal, animal focus, cartoon costume, bipedal, dog paw, 3 fingers, kemono, beautiful eyes, highres, tuft, white belly, thick thighs,
Recommended negative prompt:
human, humanoid, name tag, easynegative, human body , breasts, horns, boring_e621_v4, By bad artist -neg, signature, watermark, necktie, rubber, rubber suit, latex bodysuit, latex,
(for scalies): animal ears, dog ears, cat ears, dog, cat
Description
FAQ
Comments (7)
70% works but hand is bad same as ordinary AI models.
Anyone knows how it solve?
could be based on the model you trained off of.
Did you train off a full-fine tuned furry base model like Fluffusion or FurryRock?
If you did already train off a furry model like that, it could also be overfitting of the model while training.
@bread001 Sure, I will train with another model soon.
trial and found:
Base model change
Anylora -> indigo : Pretty Good
indigo -> Anylora : BAD
Anylora is best for multiple model use.
Note:
At first, I captioned with WD tagger and trained with anylora, but the training failed because the amount of information in the captions that WD tagger gives is too small.
Second, I captioned with ML-Danbooru tagger and trained with anylora, it's pretty good.