Pony Krea2 (WD14)
This Image Shows Both Loras vs Base Image
Trained on over 5k images from the top 5 onsite pony models
Makes 1girl far more safe for use
Over 50k WD14 triggers have been trained in Krea2
Low Rank
As suggested by @retr0
A 14 megabyte Lora with a powerful punch. This Lora learned and retained most of the WD14 concepts but does not effect style in the same way the full rank version does. So anime, photo, drawing, sketch, should work better with this version.
Description
FAQ
Comments (12)
but why
The amount of concepts I was able to add to Krea2 using this method would likely take 10s of or 100s of thousands of hours to reinforce with natural language
Also base Krea2 can do 2k images but as an example this image would only be possible with the training - https://civitai.red/images/137682502
@GlowingGuardianGirl Basic WD14 concepts where reinforced - I have not tested seed to seed vulva size
is 5k images to low for danbooru tags?
Depends on the concept. Krea2 seems to have been trained on a very wide corpus possible with Civitai images scraped
So full lora lets you use tags without affecting the style? is that what it does?
That would be low_rank version see the image for reference https://civitai.red/images/137695082
@Felldude Is there any benefit on using the full lora?
@sovietlat The results are closer to realism with the full lora - would depend on your goal