An attempt at a domain lora introducing more furry info for Anima.
As of version 7 some artist data is now baked into the lora (see below), if you want a specific artist next version just let me know with a comment.
Weights from .7 up to 1 seem stable.
Hiresfix or something similar will work wonders, this model was trained at a high resolution.
Feral anatomy is still highly unstable, it is best to anchor it with both "feral" and "quadruped" tags for now.
For Biped/Anthro anatomy just use the tag "anthro".
Versions 1-3 are trained on Anima preview 2, version 4 is using Anima Preview 3. Version 5-7 are trained on Anima Base.
Useful tag advise:
- Score tags in the positive prompt can cause human features to leak through. if you're having issues you can try to remove them.
- If you want to tag for a specific animal you usually just need that animal's simple species tag, for example dog, mouse, horse, bunny.
- Some useful trigger tags include: "canine pussy", "horizontal cloaca", "knotted penis", "feral + quadruped for feral characters", "anthro and sometimes biped for non feral".
- Try using this on base Anima. Yes I mean actual base Anima.
_________________________________________________________________________
the model now contains some artist data, it's just a start but give them a go!
@azre, @backsash, @butterchalk, @cobaltSnow, @dagasi, @dodudoru, @einshelm, @ellael, @etheross, @etoya, @evomanaphy, @frezezyk, @jingzhou14848, @kojisauce, @lodetail, @nezumi, @pekopon, @photonoko, @porldraws, @rabbitadvisory, @sepiruth, @sicmop, @vksuika, @whisperfoot, @xennos, @yuio, @zinfyu, @zizujel,
_________________________________________________________________________
If you want to merge this into a model go ahead! Just throw me a note or something.
Happy Generating!
Description
Updated and trained on anime 1.0.
Images are trained at 1536 now to match anima base and generations look less muddy.
Main issue i've noticed is a bias toward fur colors on species, which can be worked with by using the negative prompt.
FAQ
Comments (3)
What was the image count and total steps+batch size on this? Just curious, doing a big artist lora myself.
~4000 images and ~44000 steps. My images are in many separate datasets and they have different repeats to attempt balancing the data.
I trained it in a few steps. The first step was epochs 1-4, I trained at lr .00008 and batch count 2.
Step 2 was epochs 4-7, lr .00005 and batch 1 to increase sharpness.
Step 3 was epochs 7-10, lr .00003 and batch 1 but added network dropout of .03 (Not a ton but it helps with over-training this late into this big of a lora.)
@Vennin what did you use to train it