This is a retrainned model based on https://civarchive.com/models/330193?modelVersionId=369981 data with additional data and cogVLM re-caption and some manual color tag adjustment. Currently is testing only
USE the recommended Lora art-full or any other cinematic Lora to counter-balance this one. It makes results significantly better. My Lora should be set at -1 and the art-full Lora at +1.
Use at -1 for better images. It literally obeys your prompt words better (to an extent).
You'll see in the example images that base PonyXL doesn't have equal prompt understanding to this Lora set on negative.
Use at 1 or more for extreme ugliness.
There was no negative prompt in all the example images.
This model has a stronger corrective measure for NSFW generation, but shouldn't have any negative effect on SFW
Description
Put <lora:Uglystyle11a:-1>. Might have to play with weight. Not tested thoroughly yet.
FAQ
Comments (6)
The problem with quality Loras and Embeddings is that 'quality' appears to be subjective. To me, the pony example without the Lora is less ugly than with it.
Yes, something like this are always highly subjective. Without thousands of experiment and comparison there's no way to even get an idea whether something positive is there. Which is why we post it so more people can try it and compare.
In addition, while it is indeed challenging to get an idea. It also doesn't necessary hurt. SD Creation can be about artistic choices, we don't always need objective measure to do something. Often time. "This feel good to use" is valid enough.
@maydaysunny While I agree that the generation of images depends on artistic choices, the "artistic" part is not really some kind of black box. Art theory is a well-defined discipline. Quality in art can be measured objectively.
I worded it poorly in my previous comment, what I should have said is that many people confuse style with quality. To them, certain styles have lower quality than other styles. For example, many people who compare cartoon art to masterpiece art would insist that the cartoon art has inferior quality, when in fact it's just a matter of style. No wonder the quality Loras learn "style" instead of "quality". For example, the preview ponies have roughly the same quality, but different styles. Flat shading + bold lines + cartoon -> pencil lines + aquarelle + less cartoon. I doubt that Loras can really learn the concept of quality. What they can learn is quality in a particular style, by training a separate quality Lora for each general art style.
@daxivi6026 The point of this Lora and my Lora aren't about the style. We know that the style changes with the Lora, that's not the main focus, since we need more variant styles in the ugly/disfigured dataset, which is only around 1600 images, mostly weighted towards fixing nsfw and humans and there are too many realistic images in the dataset that are similar.
The point is to have a large enough dataset of failures to create greater obedience and better image cohesion.
Twilight Sparkles star on her butt looks more accurate to the original symbol with the Lora than without. That's the main point of us testing these Lora's.
It's not about style atm. There has been weird behaviour with my ugly_style Lora as well, where it followed the nsfw prompt better with the different Lora at -1.
We actually need a larger dataset with more instances of failure and blurriness with many different styles.
It's not about the style or artistic merit, it's about removing failures through negating them statistically. The problem is that I, as one person, have generated failures according to my bias.
We want structural failure taught in the dataset. Not style. Which is why we need more people generating f*ckups in different styles.
I might even go specifically for human hands and human feet, as well as faces specifically. Even with this small crappy dataset, I have seen signs of emergent properties, such as following the prompt better, in the ugly_style Lora.
It mostly repairs nsfw, human images specifically, because that's what the anti-Lora is trained on.
It's not about artistic style. We are testing if it works at all, with extremely unrefined datasets.
@daxivi6026 No prolbm at all, perfectly understandable. Just clipping in to ensure that we understand the scientific side of things, however, the nature of this type of lora is challenging to objectively measure. As for the name, it is just that. A name.
I think maybe you push too hard,this lora is making the subject more prominent too much，even i had any erotic words in prompt, and the whole picture become overfitting.
In the absence of more data or someone who can adjudicate between style and features.Maybe your thought is right,But right this time,in the face of insufficient computing power and developer capability, various factors that make generating images this convenient cannot be separated and dealt with.
Well,maybe one day,we will see.
English is not my native language,I use GPT to translate,I am not sure if you can understand the translated text.
Anyway good luck!!!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.