This is a true slider lora so it should not change the image to much compared to normal loras. Its also no problem combining this lora with other loras.
The lora is very weak, so if the prompt specify "high-waist" etc, the prompt will probably override most of the lora-effect.
I RECOMMEND COMBINING medium and strong for best results.
Strong have a tendency to unbutton or rip the pants. But with medium at around 4.0-5.0 combined with strong at around 1.0-1.5 its perfect imo
medium is has better quality, but strong works where medium is limited.
Recommended range
medium: -5 to 2 (negative = lower)
strong: -4 to 4 (negative = lower)
All my work is 100% free. The only thing that really matters to me is people leaving a review, good or bad. I will NEVER remove or block anyone for leaving a bad review, so don't be afraid to give me criticism.
I'm stating this because when a creator blocks someone for leaving a bad review, that person can no longer see any of the creator's models, essentially muting their critics. This is something I'm highly critical of, and I believe it's a fundamental quality problem on this site.
Description
This is my "medium" variant, and will not pull the pants so low that it exposes anything (in most cases)
FAQ
Comments (8)
Did you train this by pure prompting, like AI Toolkit does? Or did you train this with data?
I trained it as a "Concept Slider" in ai toolkit @Jellai
So the data i used was just 6 contol images of normal people.
I have described my method in one of my other loras.
But to sumarize it short, i sample every 10 steps. And stop often to reprompt to fine-tune.
If the training goes out of my target, i restart the training instead of trying to fix it.
@Loraholic okay. well, I'm very impressed with the outcome. In my experience, using the slider training always creates a noticeable difference between the two ends, where one (always the positive) is saturated/high contrast/exaggerated, and the other side (always the negative) is desaturated/low contrast/subdued, no matter what the topic is. I do try to compensate for about 25 steps in different points in the lora training (like the tutorial says to), but it still doesn't quite get there.
But maybe that's a Z-Image Turbo thing? Maybe Klein handles it a lot better. Either way, I know it's not perfect, and I think your output looks great. You handle the compensations well.
@Jellai If the model starts to drift into overexposed etc, start over instead of trying to fix. My best slider loras are only 150-200 steps. And i suspect the best loras is when the training overshoot, but randomly hits the sweetspot when it does, and only frequent samples will catch these scenarios. In my experience slider-loras can be more of a lottery than good planning and prompting.
And btw, i use Differential Guidance from time to time. This lora was made with Differential Guidance on 3.
I have tried to train loras for Klein, but im having a hard time getting good results. I suspect Klein need a lot more steps, and i think the base data in Klein is a lot more concentrated towards the "norm". zImage on the other hand have a lot of data hidden deep inside the model, that only slider-loras can bring out.
Thanks for sharing. Will it slightly damage the facial features of the celebrity characters that are already present in the original ZIT model? (I've tried other similar sliders like LoRa, and they still cause slight damage – for example, drawLEO MESSI，USE LORA ，the similarity drops from 99% change to 95%).
From what i have tested, no.
At least not persons that are well known to the model, and within the recommended range.