The purpose of this model is to generate really ugly people.
As other people have shown, it can also be used with controlnet or img2img.
SDXL v2 should be used on +- strength 0.5
Description
FAQ
Comments (17)
Why do they all look like Smeagol?
Dataset just containing images of smeagol🤫
I love your Lora !!!
How did you train this? What was the input data? :) It's amazing!
input data for v1 was dalle 3 and for v2 its midjourney. no real life images were included due to obvious reasons :)
family album?
@zwaetschge Yeah yeah, but I meant more like what kind of images, how many, resolution and settings. I mean the file is TINY and the results are awesome :)
@32Bitshifter yes. glad you recognized your family without any hint
@kallamamran all images are about 1024x1024. Some 3:4 and some 16:9. And it was about 50images. About the size, so far all my flux loras trained on civitai itself are below 20mb, which is amazing compared to the ~150mb per sdxl lora.
@zwaetschge Do you mind shaing your settings? You gotta use very low ranks right?
@kallamamran theres not much to share as i really use the standard settings of civitais trainer. I raise the epochs from 5 to 20 so i get about 1000 steps at least and then i download step 5,10,15 and 20. I check which of those works the best and if they all suck/are a bit off i either check the inbetween epochs or retrain with new tags/dataset. Last case happened sometimes with sdxl, but for flux most of the time 10 epochs are perfectly fine for me, while sometimes 20 epochs are needed to get the model to learn a concept properly. As example my chrome lora is fine with 10 epochs. The 20 epoch one does work too, but orients itself too much at the original dataset.
@zwaetschge Ah, I see. Thanks for replying. I have not used the CivitAI trainer since I can train for free (without using buzz) at home. I have been using AI-Toolkit, so it might be a little different 😊 Would you mind showing a few of your dataset images.
@kallamamran ah yea understandable. i mainly use the civitai trainer for parallel training of 5 models the same and it doesnt block my 3060 all the time :D
Thanks fore the nightmares. Endorsing.
God thinks there is a difference in between ugly and caricatures. Some people would simply like to generate normal average ugly looking people because created all levels of beauty.
well the lora in the examples is set to 1 to show what it can do. if you only want a bit of uglyness, set it lower.
for flux at least, this is legit one of the best Loras on here (low strength). many regards. i think it's almost imperative you make a schnell version(!).. please and thanks again