Support me at https://ko-fi.com/elderagent I'm broke!
I only post on civit and tensor.art do not upload my models without permission.
Commissioned IL LoRa of Priya Mangal from Turning Red
Prompts:
priyatr, dark-skinned female, wavy dark brown hair, brown eyes, black round glasses, nose piercing, yellow hoop earrings
grey wristwatch, bracelet on one hand
collared yellow striped shirt, brown cargo pants, white sneakers
white shirt with green edges, green shorts, white sneakers
half-closed eyes
Description
Comments (5)
Very impressive model, has the 3D style without needing to tag or reinforce it, and far more flexible than i expected it to be outfit and hair wise.
The only small issue i really had was with the nose piercing, occasionally getting a mutated nose as my checkpoint struggled to decide what part of the nose was pierced. I managed to reinforce it by using 'bridge piercing' alongside the 'nose piercing' tag which stabilized it fairly well.
Was this trained on-site if you don't mind me asking? I'm curious what settings you use to get such good results out of AI Toolkit.
Thanks for the great lora! (b ᵔ▽ᵔ)b
Yeah all my models are trained on site, when using the Toolkit it's always a coin toss, it isn't strange that it takes several training attempts to get good results
@elderagent I've heard that a lot, there's been more than a few creators recently posting articles about how they struggle with AI Toolkit compared to Kohya, having poor results in direct comparison.
I would argue that it IS strange that it takes several attempts to get decent results considering it's meant to be better tbh. Though i admit i haven't much experience with lora creation beyond a few guides and chats about it with other creators.
I was hoping that you might have some insight on the settings best to use, and to find out if you were willing to share them with other creators, or at least discuss your experiences with AI Toolkit vs Kohya with them too if that's not too strange a request?
@NeilSylver Yeah I posted an article a while ago and it's still pretty much on point, the only variable is whatever process the toolkit follows, other than having a good quality dataset
@elderagent So settings-wise you haven't pinpointed a good configuration that consistently works for Illustrious models? I know that NanashiAnon has experimented and compared results between Kohya & AI Toolkit with Kohya having favorable results even on 'failed' lora, and p4ch1nk0 dislikes it so much that they've made a contest to draw attention to the issue.
If you ever have the time, would you be willing to at least talk to them and others in the comments section about your settings or thoughts on the matter in that article with them? https://civitai.red/articles/35450/savekohya Sorry if i'm being a little annoying about this but the more discussion there is about this the better imo, and even if Civit doesn't care enough to keep Kohya then you could at least share configuration settings with each other and find out what works and what doesn't for Illusrious lora :)