V2 is a Slider LoRA and will retain the original image look.
V1 is not a Slider LoRA and will alter the image.
It supports a wide range from -10 to 15, with best results typically below 10.
Because a slider only amplifies what the base model already knows, it’s less extreme than V1—but delivers significantly higher quality.
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 was originaly just a test, but it turned out quite good imo. So i released it.
This (V1) was trained on a bad dataset so it messes with faces and the general quality of the output.
But it has some usecases where other sollutions dont work.
FAQ
Comments (14)
Wow thank you so much for this lora every wedgie prompt would have the underwear floating. Youre the best
Thank you. Happy you liked it.
May I asked what tools/config you used to train the LoRa?
I tired training my own character LoRa and the result impacts the image quality way too much, while not effecting the image at all if under 0.7 strenght.
This one, however, doesn't seems to destroy image quality that much from the quick tests I did.
I used “AI Toolkit” (by Ostris on GitHub).
I think this is the only way to properly train Z-LoRAs for now, since the turbo model is the only model available. AI Toolkit has a adapter that don’t fry the model when training on a turbo model.
The only settings I changed from the stock Z-image template were:
- Quantization of the model to float8 (because of 12 GB VRAM)
- Differential Guidance activated at 3
- Limited training to only 1024×1024
- High noise for the first 500 or 750 steps, balanced for the rest (don’t remember exactly)
Used 11 images at 1024×1024 as the dataset. All relatively close-up without face. Different colors and positions, etc.
Make cool innie pussy, hehe.
But one face for all gens :(
The model is trained without any faces. But training will always fry a lot of the base model.
If you use nationality, and spesify the character more, you should be able to make different faces.
And i think it turned out way better then i would have ever thought it would.
I am with you on that, we live in very interesting times!
Was this trained using SDXL AI images and not real people? Unfortunately this does ruin z-image's varied and realistic faces and body proportions.
I stated this in my description that it removes some of the details that make z-image so good. Images was a mix from different models in my image-collection. But mainly SDXL.
Because it originaly was just a test, i did not really think much about the quality of the source images, and instead focused on mixing up positions and colors.
I will make a V2, probably in a day or two. I have focused on a much more demanding concept since this lora, and have not had the time to make a V2 of this.
Thank you for the feedback
Any tips for zturbo training parameters would be greatly appreciated if you have time to share and explain.
You say it is outdated and there are better options, but on Civitai I don't find any alternatives for Z-Image.
Now you have another alternative! V2 is out. Hope you like it
@Loraholic will it work on turbo?
@alex9692 All sample images are made on turbo.
Base is more for making loras or difficult compositions that need negative prompt.
Turbo has the best quality.