If you enjoy my content, consider supporting me on Ko-Fi! ☕
Requests are always welcome in my channel in the Furry Diffusion discord! 📝
--v1 NXL-- [LoCon]
Trained & tested on NoobAI EPS v1.1.
Updated for Noob, reworked the dataset while tweaking tags.
Recommended positives: mhkulv, goldmantle
--v2 XL--
Trained and tested on PonyXL v6.
Modified dataset again, changed activation tags. Should be more flexible now, as Pony's original interpretation of "kulve taroth" was a bit damaging.
Recommended positives: mhkulv, goldcloak
--v1 XL--
Trained on PonyXL v6, tested on AutismMix Pony.
Modified dataset w/ slightly tweaked tags.
Recommended positives: kulve taroth, gold cape*
*Slightly finicky, moreso on ferals.
--v3--
Trained on fluffyrock-1088-zsnr-vpred-e257-minsnr-e31-ema, tested on EasyFluff v11.2.
This was trained with V-Prediction & may produce less consistent results when used with older models that lack VPred support.
A smaller update to add vpred support, with a revised dataset. Tails are a better now, and ferals can now be generated, though not perfectly.
General positives: kulve taroth, grey scales, goldcloak*,
*For the golden "cape"
--v2--
Trained on fluffyrock-576-704-832-960-1088-lion-low-lr-e83-offset-noise-e68 and tested on yiffymix 3.1.
Tested at weight of 1.
Recommended positives: kulve taroth, dark grey body, goldcloak*, tail
*For the golden "cape"
--v1--
Trained on Fluffyrock, tested on Yiffymix 3.1.
Recommended positives: kulve taroth, gray scales, gold accents, scalie, tailDescription
FAQ
Comments (5)
Nice update to the Kulve Lora! If you don't mind me asking, what were the training stats you used for the v2 model?
The settings used here were still a bit experimental (with a non-standard optimizer), but the general gist of it was a batch size of 8 w/ a gradient accumulation of 2. TE & Unet LRs were 0.00015 & 0.0004 respectively. Still tuning these new settings a little, but I'll be adding them in detail to my guide soon-ish. :)
@Valstrix Nice! Thank you. I've been trying to make a bunch of MH Loras of my own as of late, thanks to your guide, with some turning out fairly well, and others not so much. Idk if it makes much difference using CivitAi's Lora making tools, but it's difficult to land the same quality as you haha. I make sure to get a pretty decent amount of images, all the highest quality I can find, but in most cases the eyes suffer the most where they're either dots or missing entirely. If you have any advice, I'd appreciate it, but otherwise I can still wait for more updates to your guide.
@TactiletheDilo Eyes are always tricky, it really depends on how clear they are in your dataset. I've come to learn that with certain datasets, it's easier to just inpaint or use adetailer on the eyes during generation rather than try and fiddle with the data repeatedly, though I usually fiddle a lot anyways lmao
As for the Civit trainer, it's perfectly fine to use (it uses Kohya as a backend, the same as me), but it obviously isn't quite as flexible as local training can be as I think some of the more advanced values are inaccessible. With my own methods I could never use it, purely because I experiment often and usually run a bunch of trainings per model (If you look at the filenames, the "aX_Y" section refers to version and training number) - I would drain my buzz rapidly with all the tweaks I like to do haha
Any chance for Somnacanth?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.