Lora trained in re-create the Dozamura's art style
I rlly like see your reviews and generations. They are helpful for fix and train better models. Share some hearts at the previews that you like and feel free to leave your comments or requests, but I can't promise anything xd.
Version 1
I found a way to reduce the LoKrs training from 14-18 hours to 4~5 and the result looks great, I'll make some more tests.
Weight: .8 - 1
Description
FAQ
Comments (12)
"Hmm... Why does this style look familiar?"
moves to the second set of previews
"Oh... the horse..."
HAHA, rlly looks like? I'm not sure if I should keep the Miku preview or the second one, just I cannot see a great difference between styles
@Konoko I meant the 4th image in currently is when I recognized the style cause it reminded me of another character they drew lol.
It's perfect! excellent!
What is "LoKrs", and why does it take so long? It does look good.
It's a type of Lycoris, it need trains more layers for that it takes a more time, you can see the details and a comparative here: https://github.com/KohakuBlueleaf/LyCORIS
@Konoko Oh, very nice. Do you know how it compares to DORA? I thought DORA was supposed to be the new hot shit, but I don't see many out there. Also, was there a common mistake you made that causted it to take so much more time? I think I'll try it out, and I don't want to make the same mistake. Thanks for the info!
@Jellai I rlly don't know. I have never tried train a DoRA. I asked to some trainers here but they said that looks like a weird piece of ... the same (?. But I'll try it for share you my toughts. About my mistakes, was Prodigy configuration and the data-set quality, you can check the configs in the meta, about the data set, I just limit it to use the images with bigger characters
@Konoko Thank you!
@Jellai Hi Jellai, sorry for the late, I just tried for first time the DoRA. Actually I'm not sure if is better than a LoKR. But in styles, could be worst or better depending of the propouse. Why?, Is because it blends better with other styles or indeed the native model's style. If you want preserve exactly how it seems, this behavior it's not the better. LoKR overwrites all, and let's empty everithing else, (for me this is the better behavior in the most style cases, this allows me to create styles based in doujins). Anyway I'll make more tests before give a final veredict xd.
@Konoko Interesting. Sounds like it's terrible for styles, but might be good for characters that you want to translate into realism, or another style? I wonder how it would handle something like actions. I imagine that's where it would be strongest based on your description.
Thanks for getting back to me on this!
@Jellai It's pretty weird, the Dora is just a Lycoris (Locon, Lokr, Loha) with the option "Weight decomposition" enabled. I tried first the Locon version (because is the one used in the model that I took as reference), I'm training a LoKr version to see how it's going
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.