IF LoRA
The default download is 3000, do not ask me why.
A general style LoRA, the main goal is to improve the realism and look of model output. The sytle is focused on modern social media and influencer aesthetics, but is not limited to that. The training data was hand picked to enhance the aesthetic quality.
Training info: Trained with AIToolkit using batch size 9, which means that the(effective steps are higher than the raw numbers. This is just how AIToolkit does it.
I could not decide which version is best, so I publish these 3_
3000 steps
6000 steps
9000 steps
Recommended weight: 0.9–2.0 The 3000 steps version may benefit from stronger weights (closer to 1.5–2.0), though results can vary.
The training data for my GP lora was included by mistake, so it can do that to now.
Description
FAQ
Comments (9)
It loves to put stupid little tattoos, logos on clothes. Oh, and watermarks 🤦♂️
Decent style but annoying for general use
Can I ask you for a prompt/example. I don't see it as often as you do, i wonder where it is coming from.
@generatorofthingsreal But you see it... I'm not prompting anything special.
in fact, I have negative prompts enabled, and "watermark, text" etc have minimal effect. I probably get it 90% of the time.
@pogo that is very strange, because I only get them on rare occasions. Are you prompting for something particular? Could be that some subset of the images have watermarks, maybe the NSFW part. Might have to try to remove those for the next run.
@generatorofthingsreal idk why we're even talking about prompt when we have established that there are images in the dataset with watermarks... there's your culprit.
Must be quite a significant amount with how strong the effect is.
Maybe you can make a workflow for mass erasing the watermarks from the dataset. Or cropping the images by just enough pixels.
The style is great, by the way. Just flawed.
@pogo We are talking about prompts because I can not reproduce your results, so I am wondering what your culprit is. Why do you see it so much more than me. That is why I ask.
@generatorofthingsreal "your culprit"... Ah, my autistic brain is freaking out. We have already solved the puzzle and you can fix it in the next go around if you want.
There are many images in the dataset from 4k nude model outlets like zishy or something and you left all the watermarks in. So when you generate full body, professional style modeling images, you get more watermarks. If you keep generating SFW images exactly like your examples, of course you won't see watermarks.
I love this LoRA. Great style. Watermark can be worked around if I really wanted. thanks for sharing it
@pogo There never was a puzzle. Just a refusal to say what you were prompting for. Your comment have made the situation clearer, you get the watermarks when you prompt for NSFW content. I checked the data, there are very few images with watermarks and most of them are not in the category you describe.
I think you are making assumptions that are wrong. Would be great to know an example prompt, so that I could test it. I really do not get that many watermarks, nsfw or sfw.
@generatorofthingsreal I'm going to tie a bow on this before I go insane. If you aren't trying to train watermarks, you shouldn't have watermarked images in the dataset.
That's the problem. Don't need to go chasing your tail for prompts when prompts aren't the problem.
It's a perfectly good LoRA. It just has that one flaw.
I will definitely be posting images with this LoRA later today. You can look at the prompts and drive yourself crazy if you really want.
Thanks.