Support the Freedom at https://ko-fi.com/lorian0x7
I'm thrilled to present the First Realistic Freedom 3, a SD3 fine-tune, the first trying to address the censorship, a step beyond the base, though not flawless. It serves as an excellent starting point for future endeavors. Hovever, there's still a lots of work to do.
With this first version, I wanted to resolve the biggest flaw of SD3, Female Anathomy, which was censored on porpouse, like if it's a Taboo. Now this model is a lot more uncensored, and it's a lot better than the base model despite not being perfect yet.
I'll try to improve this version in the coming weeks
If you value freedom of open-source AI, consider supporting me on Ko-fi.
https://ko-fi.com/lorian0x7
recommanded settings:
Euler - CFG 3-6
It can be hit and miss sometimes
Description
First NSFW training ever done for SD3
FAQ
Comments (16)
It's really great~This is an exciting result.
Interesting development actually managing to reverse SD3's lobotomy. Congratulations to that!
How is the file size so large though? The regular SD3 medium had less than 5GB iirc.
Makes beautiful landscapes, but if you're hoping to make NSFW images you're going to be pretty disappointed compared to current SDXL and Pony finetunes. The fact ADetailer and Kohya Hires doesn't work with SD3 is also a huge downside, although that of course is not the model's creator fault. All in all an interesting experiment.
bravo~~~~~!!!!
Boobs confirmed with each request (with varying degrees of quality). sitting is a challenge without spaghetti legs. But as a first itiration, great. now we need about 20 more models merging and working before we hit XL levels of course...but hey, one small step at a time in the right direction. nice work. Worth it for the text.
I tried using ComfyUI with the settings Euler - CFG 3-6, but the resulting images were very blurry. Can you share a workflow
But does it do penises?
initial sd3 render very realistic .. why this is plastic ?!
sorry , this absolute garbage .
Was the dataset trained on images with tag captions? NLP being the very one thing that SD3 is good at?
I was going to complain about the behavior/limitations I saw, but most of them aren't actually yours. They're in the official SD3 checkpoint model. (See my comment there.)
You guys have to keep in mind that the original SD3 model is trained with multiple text encoders, which makes the model not as good as it could've been.
It also has a lot more channels, and more channels means more training, which is a challenge by itself.
On top of which, SAI scrambled the anatomy which is undoubtedly going to be an endeavor for the community to restore.
If you thought XL was developing slow compared to 1.5, SD3 will accelerate even slower than XL.
Until we get the small model of SD3, I suggest everyone to just have some patience as training on the small model is going to be a lot easier, and a lot more people gonna be able to run those models.
If my vote counts for anything, I vote to boycott SD3-Medium like everyone did with SD2.
How? How are you generating something with this model? on cfg 3-6 everything is oversaturated 1 and smaller are only usable... anatomy also wrong 5 tries of making a young woman and all with incorrect anatomy
does this model include CLIP or t5xxl? I am sorry but I don't like models that I have to load external text encoders.
Is there anything in this model other than photorealistic white skinny young female human nudity?
Any plans for a 3.5 model? mad potential!
Details
Files
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.