A fully original LoRA to impart a high quality epic fantasy oil painting look to your Illustrious images. Dataset was crafted specifically to train this one, so hope you enjoy it!
Trigger: painting_il
Description
FAQ
Comments (5)
Hey! The provided results look nice... but the colors in some are overburned.
This is not the first time I have seen this in IL checkpoints and LORAs, and I'm interested if that was the case with training material or if this is the IL's behavior itself?
Honestly, its probably because of my checkpoint I used, which is my own attempt at a merge; I liked the result despite it for stylistic reason (I tried it on my usual, better IL checkpoints), so definitely try it out on your favorites and it likely won't happen.
@Alienor_ Noticed problems with face inpainting: even on 0.4 denoise and 0.6 value of this LORA it renders very out-of-style faces... but maybe that is a problem with a mix with my chosen checkpoint.
I use 1024x1024, 0.4-0.6 denoise for faces in aDetailer.
But hey! The hand that holds an axe took only 2 times to re-render (it was around 40 tries without this LORA) :D
@Damacy Yeah the ADetailer issue I've had, and I've had it on multiple checkpoints, even before making this LoRA. Wondering if its something in some of the IL models though; that's not specific to this LoRA.
(I use 512x512 and 0.4 though!)
Which checkpoint were you using?
@Alienor_ Aha... Maybe that is the case with semi-realistic and Western-like styles in Illustrious and NoobAI.
I played with NTR MIX and MonsterCoffeeC4 for a bit, but as I'm not a huge fan of flat shade animeh, I used a few authors: Sakimichan, Kittew, bbc-chan, etc. Had no problems with upscaling such faces.
But then I switched to Smooth Mix, and got a two-sided story: anime-style faces upscale as usual, but semi and western will go way over the top with changes and details on a low denoise like 0.4 (to get a similar effect in PonyXL you need ~0.6)
Hmm... I guess we can try really-really low values like 0.2 and maybe 512x512 to avoid severe changes :\