This is a work-in-progress LoRA for Pixart Sigma featuring two concepts that seem to scare other models: Nude women, and women laying in grass!
This is an alpha, trained on 350 images for 45 epochs (anything beyond was unusably distorted). I am going to compile another dataset to further refine it, but I wanted to get this out as a proof of concept.
The model appears to work better with the word "nude" instead of "naked"
Description
Initial release
FAQ
Comments (19)
You know that SD3 will get finetuned soon... Right?
I hope so, however finetuners have spoken about how SD3 is nowhere near as easy to finetune as expected, plus licensing for SD3 and Cascade has scared off many of the larger groups that use finetunes since it affects the ability to recoup costs.
The developer of Pony in particular has struggled to get any clarification on getting enterprise licensing, to the point that they will be focusing instead on releasing a v6.9 on SDXL.
This is really just to get more attention to some of the other amazing models that deserve more love! I’ll also make one for Lumina-t2i when a LoRA GUI for it comes out!
@Thatrandomjew That's the thing. I struggle to see how $20 a month is a big deal if you've already made $1 million with the model. I think it's fair that people who benefit the most from using the model at least pay the bare minimum. SD3 will still evolve without the large finetuners. It will just be painfully long process instead. Hope you get success with your projects tho. The more the merrier!
@Dazrock the $20 a month isn’t the issue. The issue is the 6000 images per month limit. Finetuners make back the money they spend on database curation, captioning, GPU runtime, etc by having online generators (like Civitai, discord bots, Runpod, etc) pay them. 6000 is maybe a week’s worth of images for a moderately sized discord bot.
They have essentially told them “you can make the finetunes, but you can’t recoup your costs”. I think that’s going to be the final nail in the coffin for SAI
@Thatrandomjew well, they said: when you need more, please contact us and we'll find a suiting deal for your enterprise license.
@StableLlama True, but the screenshots of how the developer of Pony was treated for trying to do exactly that speak for themselves. I'm not gonna go too far into what was said, other than it was entirely unprofessional, completely unconscionable, and is the reason the new focus is on Pony 6.9 and not on a SD3 based model
Apparently problems are occurring with fine tunes on SD 3. They'll see if they can fix it.
@Meowser The current info indicates that they found the part of the weights that pertain to NSFW content and nuked it.
Also people notice that putting NSFW terms like breasts, etc actually result in better results, almost like they deliberately took bad images and miscaptioned it with NSFW terms in order to poison the model
becoming doubtful
Probably not. Why waste time fine-tuning a model when the new license means that if you share it they can pursue you to have it removed because they own all derivatives. They abandoned open-source, so open-source is going to abandon them back. They can produce their own tools and fine-tunes and we'll just move on to greener pastures and they can sink or swim.
Nice try! Pixart is not overly censored and can even produce revealing costumes and nudity!
Even without a finetune it was already capable of boobs
This is so unsafe my cat sat on me and broke my legs. Edit: Don't look at my profile it's unsafe.
lmao gottem, anything to make a stab at SAI is a good thing in my eyes. Good job!
Do you have the VRAM do finetune instead of just making a Lora?
VRAM is no issue, it can train on 12gb. I only have 16gb of RAM though which is an issue. I'm running a Legion Go with an external 4070ti which limits what I can do. If I could afford it, I would be buying the new Ally X with 24gb or a gaming laptop, but money amirite?
That being said, I am trying to see what I can do to make a proper NSFW checkpoint based on SIgma 900m
@Thatrandomjew You could probably finetune it at batch size 8 under 16gb. I'm currently doing it at batch size 16 and it's using less than 20gb if I recall right. You'll probably have much better result with finetune than Lora. From my experience so far, sigma is really undertrained for anything waist down. The training data looks like it was mostly close up portraits and headshots.
@2BlackChicken So I have good news and bad news:
The bad news is that I cannot get a working finetune saved in Onetrainer
The good news is that I have made a new NSFW finetune based off of the 600m and 900m models and have backups from right before they converged.
I just have zero clue as to why the images come out distorted and unusable in comfyui and sd.next while the samples in Onetrainer look perfect