About IL V3: The first versions were created using a dataset of real photographs, which had its consequences. Version 3 was created using a dataset that was a nearly 50/50 mix of real photos and the best artwork made thanks to the first versions. Initially, I tried making the entire dataset solely from artwork, but the results were too cartoonish and didn't mesh well with newer styles. So now I'm trying to make all dataset for LoRA 50%/50% real photos \ artwork.
iLLustrious V3:
face in ass, iLLustrious V1:
face in ass, femdom,grabbed him by hair,Pony V1:
face in ass,Description
FAQ
Comments (4)
I made some tests (Illustrious V2) and it plays pretty nice with the prompt.
3 consecutive txt2img generations (seed growing +1 per image) with very simple prompt can be viewed in this post: https://civitai.com/posts/13474509
Sadly I forgot to prompt for shoes on the poor dude, so he is wearing high heels. But the point is the generation adheres to the prompt very well.
For me, even as low as 0.7 strength, it makes the characters realistic which is yucky.
Same about the realism that this Lora introduced. All of your models do it. Any advice on how to circumvent that?
My rating for version v3 [IL] is 7/10. It's the highest rating so far for this LORA series.
I tested them in Automatic 1111 and got good results.
Thank you for sharing and for your work.