Gives creamy skin and warm dark brassy tones, similar to the bdsm comic artist Ferres.
Best around strength 0.8. "forest,trees" in the negative can help give different backgrounds.
Using some of the popular style loras alongside it really helps the details.
dpm_2_ancestral works well.
Note: When opened with the remix button the lora strengths aren't set properly, check the prompt.
Please use responsibly.
Description
FAQ
Comments (6)
Hey, great job! I've been trying to make Ferres style for a while, but never managed to do it right... well, was rather inexperienced back then.
I see from the example images that you mostly used his "middle" works - nothing from the newest or oldest works of his, right? Jinni, Tempest, Masque, etc?
Nice!
I too have been trying for sometime.
Yes this only used 28 images from just a couple three comics, all with the light skin/dark background contrast and simple composition.
I concentrated on removing images that were perhaps giving it bad or mixed ideas, such as a lattice work background that it would REALLY pick up on.
Previously tried with many more images, but it would loose the nice contrast, muddy the style and induce weird anatomy.
Perhaps there's something that can de done with the captioning to help, as It would be nice to allow it access to more variation.
(Used the onsite trainer, 4 batch size, 8 repeats and with the flipAugmentation The captioning is mostly untouched wd14 with about 8 normal written captions. I added a trigger word but it works better without it.)
@TrinityDeVayne thank you for such a detailed answer! And for posting the training settings as well!! Only 28 images in dataset? So few? How many steps did it take?
Previously I tried to use 40-60 images, it didn't work well on SD 1.5. Well... maybe it did - I just didn't put any preview prompts, got weirded out by the mess Civitai generated for me, and never used the results I got. Who knows, maybe it did work out after all? Currently I am slowly gathering my new datset - about 700 images already. Slowly clean it up from speech bubbles, watermarks, and image frames. Some I upscale. I plant to make several variants of it - one with a big dataset with every image(or the best ones) and several more with images from different comics. Mostly his earlier works because I don't really like Ferres' new stuff - square faces, dead eyes, disproportions, etc.
I too use onsite generator because my PC isn't strong enough. Well, I haven't tested it, but I'm pretty sure that my 1050 ti will struggle with Lora training.
@JxC The steps were like 448 but with a batch of 4, so really 1792, I've always found there is no point in going much over 2000.
Certainly recommend testing it on your own computer if you can, I think the civit preview images aren't a good representation. The efficiency nodes for comfy ui have a great x/y plot option.
Good luck with your big dataset!
@TrinityDeVayne hm, good to know. I also noticed by my first character Loras that going over 2000 is wasteful in terms of buzz, besides the Lora turns out very rigid and overtrained. Now, my character Loras rarely exceed 700-900 steps and they usually turn out flexible and overall great.
Yep, even though I train on Civitai, 99% of all my generations I do on my own PC :)
And thank you! Good luck to you aswell with your future models!
can you do this for illustrious?