The beautiful Milana Vayntrub who plays Lily on the AT&T commercials. There wasn't a lora of her for Flux.
Description
My first lora generated from 45 images.
FAQ
Comments (6)
Very cool, thank you! I noticed horizontal lines/bands on most of my outputs that show up in your examples as well. Seems to happen on any strength 0.9 or up, but I need those strengths to maintain her likeness. Any info on this or why it happens?
I don't know what is causing that. It is very disappointing. I didn't see anything particularly in the training data that might have caused it. I might try some different training data and settings and see if I can get rid of it. I really like her.
@chipjohansen99361 Thanks so much for your efforts....your Lora is the best Milana lora around currently, regardless of the bands :) I've found that dropping the guidance down to 1.8-2 really gives the best results and gets away from the 'cartoonish' look and more of how she looks in real life. I'll upload a few of my better examples from your Lora later
@tl986 I did publish a V2 version of Milana that is much better than the V1 version.
@chipjohansen99361 Thanks so much for your hard work brother...I really appreciate it! I gave your v2 a shot, and it looks like the banding is indeed completely gone now! Completely personal preference, but this new v2 seems like a slightly more younger, 'idealized' version of her. It's hard to explain...like her nose and jawline, etc...all seem more anime'ish like she's in her early 20's. I tried experimenting with a whole bunch of different guidance values and pushing the weight all the way up to 1.7 which was the only way I could get a more semi-realistic mature look that represents how she looks in commercials. But then it made her face look totally airbrushed and plasticy. Your v1 really hit the spot in terms of nailing her likeness to me...I've published 2 examples of mine. The only small gripe was the banding in some of the pics. Curious....did you train this new v2 on a different photoset? I'm wondering what accounts for her slightly different apperance. Just a side note: but I really didn't have to tweak much with v1...sweet spot was a strength of 1.0-1.1 and got consistently accurate output with guidance between 2 - 3
Not a complaint at all, just some info...thanks again!
@tl986 No, it was the same training data. That was the difference between using the first epoch and the fourth epoch. Maybe the second or third epoch would have been the sweet spot.