A nude/NSFW capable model that focuses on images with a female primary subject.
It is heavily biased towards nude images.
Recommended Settings
Sampler is dpmpp_2m (DPM++ 2M) and the scheduler is beta
No VAE or CLIP is baked in. Use separate sources for those.
Each sample/preview image contains the used workflow. Here's a quick article with simpler more beginner-friendly workflow. This is a recommended starting point.
FP8 Download
For some reason, when you upload multiple formats (BF16 + FP8) on one model, which they now support, they hide the FP8 under the BF16 model group in the side menu...
Early Access is enabled to support the development of new versions and models.
Description
Improved female nudity and slight improvement to male + female simultaneous characters. But still nowhere near working.
FAQ
Comments (23)
I get this error You do not have CLIP state dict!
Someone know how to solve it?
Are you using Forge?
I've only tried it in ComfyUI myself.
you haven't put your clip models in, some models have them baked in most do not.
If you are trying to use this model's CLIP, then yeah, that won't work.
You need to load CLIP separately.
I haven't trained CLIP for this model, so there's no good reason to include it.
Better load it separately.
You need to download 3 files and put them in VAE and text_encoder folders https://github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050
5k? Bruh...
Think of it like this: It's free in 10 days. Use version 2.4 until then. I'm sure it'll serve you well enough until the time this comes out of Early Access.
Supporting here is entirely optional and I don't think less of anyone that waits for the free model. If you like it, you can give thanks by uploading a picture or adding a review.
Thanks for checking it out and caring enough to comment! I hope you like the model when it's out of Early Access! <3
My comment wasn't meant to be ironic or mocking. I was more like stunned by the amount requested for early access. I am supporting a lot of creators here. some with 100 while others with 500 or 1000 or sometimes even more. However, i considered 5k for support is way much more than usual. Anyways, i am sure it is a great model. Cheers
@6tZ ngl whatever way you look at it 5k for a single person to support is kinda ridiclous, i mean you couldve made it atleast 2k more reasonable, and people may consider it much more, 5k is just crazy high, even if its "early access and gonna be free", pretty sure more people wouldve donated the 2k and so in result more buzz in total donated, but whatever
@twinengines1158 Exactly my point
@twinengines1158 Fair enough. I mean in the end it's up to me to determine the price. I think I might try 2k for the next version to see and compare. And maybe 200 or 500 or something very small for one version.
It's hard to gague the actual values, since comparing model to model won't really work. Comparing version to version is better, but has other drawbacks.
Thanks for the feedback regardless! Much appreciated.
@6tZ over 10 days and not free yet? Or something did change?
@Dbooya What do you mean?
Version 2.5 is free, you need to switch model at the top bar.
I've added a gguf version of 2.5 which is in early access right now.
Hi.
Great checkpoint.
I used the version 1.0 fp8 from the beginning and I'm happy with it. :) But.........
I just see that updates are available, and tried version 2.4.fp8, but unfortunately, with this version I have the horizontal stripes when upscaling (known artifact made by FLUX when upscaling to 2k+. see https://github.com/lllyasviel/stable-diffusion-webui-forge/issues/1712).
I don't know what you did (or not did) between the two, but I had to revert back to version 1.0 !
I didn't tried yet the others versions (2.1-2.2-2.3), but I tried a lot of sampler and scheduler and can not get ride of them.
I can confirm that I don't have the stripes when upscaling with the version 1.0 but always have with 2.4. (tried exact same prompt with both).
Anyway, thanks a lot for your work !!
EDIT :
I just wanted to say that I have stripes on every checkpoints I tested (even the original FLUX 1D fp8).
It's more or less depending of each checkpoint, the prompt, and the result image, but I have them on all checkpoints tested.
BUT not (never) on your v1.0 !
I'm using ComfyUI with the "Load Diffusion Model" node.
EDIT 2 :
Just realized that I used it with Euler - Beta, but the quality is even better (by far) with Dpmpp-2m - Beta !! :) (even if it looks a little more "cartoonish").
Thank you so much for the detailed comment and help! The striped line curse is so annoying. I know it's a Flux setting thing, but haven't been able to figure out how I can affect it. Maybe some models are just more prone to it? Not sure at all honestly.
It's good to point out the stability of older versions. I will definitely revisit previous version. I think 2.5 is unmanageable from where it is. It's producing great images, but not listening well to the prompt at all, and it's performing much worse with LoRAs on top. So I'll definitely go back to a previous version for my continued work on the model.
I'll mess around a bit more with the sampler and scheduler next time, to see which outputs are better. My problem with the older version of the model was the Flux-like-ness, and shiny skin. This got improved over time, but maybe an older version of the model with different settings is preferrable. It's just so painful to test things on Flux as generations are so slow :)
If you have any more tips, suggestions or feedback, I'm more than all ears!
Thanks! <3
@6tZ
Thanks a lot for your answer and sincerity ;) :)
For your information, I finally managed to test all versions of your checkpoint, to verify if they have stripes when upscaling (and of course, because I would give a try a the newest version possible :) ).
v1.0 -> no stripes (in fact, seeking deep, I could see them on only 1 image on a perfectly uniform blue sky, but it was almost unnoticeable).
v2.1 -> Some stripes are visible on certain parts of almost all the generated images (usually when light or low uniform lightning). But it's very low compared to the newer versions.
v2.2, 2.3, 2.4 -> All have visible stripes on lot of images (same places as v2.1), on almost all images, if not all.
Keep in mind that this is what I have at home, but it could maybe not be representative of all other people !
So, for me, unfortunately, I had to keep the version 1.0.
About what cause the stripes, some people says that it could be caused by the vae file. I don't have sufficient knowledge to comment that, but I tried 2 different vae for FLUX - one of 327Mb and the other of 163Mb, and don't see any difference between the 2 images.
I don't have any knowledge about training checkpoint (I made some FLUX Lora for myself, but using a pre-made workflow), anyway, if I could help on any task, don't hesitate to ask.
Regards, and thanks again.
@zetofzone204 I've had some luck with turning up the Guidance setting, but if I push it any higher than 3 it starts to make any model look like porcelain and plastic -- you know, that "schnell" look, only dialed up to eleven. I've also found that adding two more steps to the scheduler (10 vs. 8 on a nominally 8 step model, or 16 vs. 14 for Dev) can be helpful, but there's not much to be gained by going past that and of course every extra step slows down generation. At least the images are unlikely to get worse if you have a couple more steps than absolutely necessary, unlike dialing up the Guidance too high which will have a negative effect on just about everything.
Checkpoint?? vae and clips are included?
No VAE and CLIP are built in.
They are best loaded separately since they weren't trained in this model.
As they haven't been modified, it would be wasteful to include them in the model.
@6tZ okay! but still great model
Where do I get them from?
@02847189727 i think huggingface
Yup, huggingface, just google for "flux vae" or "flux clip", or follow some guide on youtube on how to get flux set up and working.