Trying to keep a beautiful glitch style of smudge + oil painting + optical flow that I stumbled upon while using cnet+xinsir.
In this version, this lora is not stable and less generalizable.
Description
FAQ
Comments (9)
Absolutely love the style, and I get way too excited for styles that happen because of freak accidents, BUT I had real trouble getting it to work on other checkpoints, was this trained on Pony or a derivative checkpoint? Anyway I would LOVE to make this work, since the example pics are gorgeous.
It is based on tponynai3 V6.5, considering it happens on this model. I trained on this model, and I guess it might only perform okay on this model (considering it's not a “normal” image style lora)
Also, making this lora more visible requires some tricks: keep only this lora and trigger words in hires, because pony's prompt interferes with image generation, and the image style of this lora is in conflict with pony's score. I don't know if civitai allows this setting or not
But honestly with what's going on, it's only capturing the effect of a small portion of the target (those glitch images), which is far from my expectations
I'm a bit confused the training dataset is very similar and the lora is still underfitting ...... I'd rather it overfitted.
@omnisaa I see, thank you for the reply. I'll admit, I like the style so much I wouldn't mind taking the dataset for a spin, if you'd like me to of course. If not that's perfectly fine, I know sometimes training these things is like taking care of yer baby, aye? :D
@KarotConCarne
https://huggingface.co/akoyaki/SDXL-PONY_my_lora_store/blob/main/totallyError.zip
Don't mind at all, if you don't mind that the style of the target isn't the same as what's shown so far, I put it on hf, feel free.
Unfortunately, since hands completely crumbled, I tried to avoid learning the wrong hand by painting it. However, I incorrectly modified the original file (I thought I was making changes to the copy) so there's no way to provide a version that isn't black-censored at the moment.
@omnisaa Most importantly: much apreciated, I'll see what I can do, secondly about the hands... oh... dear... me. I think I'll have an idea what to do with the black spots, no worries, though it might need some experimentation. That's what we're here for tho, so I'm not complaining :D
@omnisaa aaand it's up:
@KarotConCarne Congratulations, it seems to have much better generalized. Unfortunately, it still seems far from the dataset...
The unique style of the dataset with its thick, blurry blocks of color on most part; the texture of the clothes mixed with glass and water make the image flow; and the girl, as if the only brilliance in the world, is illuminated. I'm fascinated by this wonderful aesthetic.
However, it seems that every lora I try only retains a little of the color blocks and color contrast, far from being a style of that.
I'm going to try train it on flux, because I have no idea how to make it on pony lol
@omnisaa No joke, yours is much more... STYLY, at the cost of working on one checkpoint, and I have no idea how to find the gold middle either :D
Good luck with training friend, I'll sure be looking forward to the flux one <3
@omnisaa Compadre... I think I have an idea, no chance on finishing today since it's gonna take me some time but... I think there's one more thing I can try to make this work <3
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.