UPDATE: v0.1 made quite bland pictures for the dieselpunk. So I fiddled with the params little bit. Now I'm satsified with the result.
It's time to take it to the next level.
This LyCORIS incorporates five *-punk genres: steampunk, valvepunk, dieselpunk, teslapunk, cogpunk (other bundles will come later)
The prompt for the preview picture grids is
<lyco:PunkBundleAI:0.9> { valvepunkai | cogpunkai | steampunkai | dieselpunkai | teslapunkai | (cogpunkai:0.6), (steampunkai:0.6) } [concept]
where [concept] is
coffee machine,
toaster,
etc.
Keep me motivated and caffeinated:
Description
FAQ
Comments (6)
Welcome to the anti-bleeding layering system. I suppose you already have noticed but the process is almost the same, just split it into folders while training and maybe decreasing a bit learning rate but most of the process can be IDEM. I felt better results with constant in LoHas. ^^
Hey, is there some reading I could study about this "anti-bleeding layering system." Like a comparison between ordinary LoRA and LoHA in context of SD? I'm absorbing theory, math and programming behind SD and everything around. It's fruitful. I see a few misconceptions circulating in related texts on Reddit and Civitai.
@konyconi same here, there are too much info but "biased" because first experiment said this... Im just experimenting with settings as I suppose you're doing. Just wondering, where are you reading about math? The first thing I did is get every setting I could, and test with the same images and started changing colors, tweaking unet, lrate, repeats, etc. I think that's the best way.
@LDWorksDavid in original papers, like https://openreview.net/pdf?id=d71n4ftoCBy https://openreview.net/pdf?id=d71n4ftoCBy
Amazing!
I do tend to prefer Loras as those can be merged into a chackpoint when needed, but having a single Lyco makes for easier changes in styles.
Thanks for this, Kony, was starting to give a bad eye at my LoRA folder, still need to trim it a little, but your bundles will help a lot, and without losing anything