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
For use in ComfyUI.
FAQ
Comments (21)
@ToAll :
It seems to be very important to use the dpmpp_2m sampler as I made some lora for myself, and they are not working well until I change the sampler from euler to dpmpp_2m !
So if you have some problems with the image quality or lora appliance, and you are using euler, try to switch to dpmpp_2m - beta.
I'm actually finding that, in fact, my selfmade lora works pretty well !
Nice! Great job and thanks for sharing!
Maybe consider writing an article about it, with some comparison images? To share the knowledge more.
Any LoRA you wanna share?
I'm still getting them w ith dpmpp_2m + beta btw, but maybe they are less frequent.
This, and the vertical thin edge artifacts are very hardcore baked into the model somewhere...
"RuntimeError: mat1 and mat2 shapes cannot be multiplied (16320x64 and 256x768)"
When using 2.4-dev-Q4_K_S-gguf with the usual Flux vae and encoder's t5xxl_fp8_e4m3fn, ae and ViT-L-14-BEST-Smooth-GMP-TE-Only-HF-format OR the standard Clip_l.
Doesn't work. Compared the exact same settings with flux1-dev-Q4_0.gguf and that works.
Your gguf is broken.
It appears to only work well in ComfyUI. I don't know what Forge is doing but it's not working well with standard GGUF conversion methods. And the built-in GGUF convertor doesn't work at all for me.
@6tZ You should ask city96, or check their discussions, the person who made the first Flux1-Dev GGUF conversion as they obviously got it right since it works on all apps not just trying to force people to install ComfyUI by putting set code here and there to prevent ForgeUI users from loading it :')
@MissBee Thanks for the tip!
Also Forge could probably add support for it ;)
@6tZ Forge does have support for GGUF, I use several GGUF models and it also works flawlessly with all of the original (first) GGUF's created. Like I said if you want to fix your conversions so they work correctly then ask city96, they have a github you can contact them on.
@MissBee Yeah, they work in Forge, if they are made there. But the tool is broken for me to create them inside of Forge. And if I create them outside of Forge, they should still be supported by Forge. It's a request for the Forge team. As ComfyUI handles them just fine.
But yes, thanks for the advice to contact them, I much appreciate it!
@6tZ It's not a request for the Forge "team" as there is no "team" it's a single person, explaining why you're having issues - lack of research - which is why I recommended reaching out to the person who made the first GGUF.
@MissBee You were right. It was not a forge issue. The problem was that the code I used to convert to gguf had a bug in it, which was fixed 5 days after I got the conversion script code.
Now with the updated code, it's all working! Thanks for the suggestions to talk to city96! They were an amazing help to figure this problem out!
All gguf-versions on this model are now updated.
If you want to you can try the new version.
<3
@6tZ Glad I could help and glad you got it resolved! ♥
@6tZ It doesn't work for me RuntimeError: Error(s) in loading state_dict for IntegratedAutoencoderKL: size mismatch for encoder.conv_out.weight: copying a param with shape torch.Size([8, 512, 3, 3]) from checkpoint, the shape in current model is torch.Size([32, 512, 3, 3]). size mismatch for encoder.conv_out.bias: copying a param with shape torch.Size([8]) from checkpoint, the shape in current model is torch.Size([32]). size mismatch for decoder.conv_in.weight: copying a param with shape torch.Size([512, 4, 3, 3]) from checkpoint, the shape in current model is torch.Size([512, 16, 3, 3]). FORGE
@dragaandroid704 Still?!
Which model version did you use? 2.4 gguf? 2.5 gguf?
Can you tell me which VAE / Text Encoders you had loaded?
Are there any Flux checkpoints yet that can actually do real hardcore NSFW or is it all just nude pinup modeling? I thought by now (it's been close to a year right?) someone would've made one. I guess Flux is just too hard to get that to work. Even with the XXX LORA, it's still just so mid.
There are no flux hardcore models.
There's just LoRAs that do a specific pose.
It tends to break down when you do multiple concepts at the same time. At least when I try it.
I think the model needs heavy brainwashing and a large training to power through it. I'm not sure if people with access to good GPUs, or larger budgets have tried this or not.
But not seeing it probably means no.
Try mine, v2 is very good and can do a lot of stuff, v3 is in the oven.
This is the link https://civitai.com/models/1295758/nsfw-flux-mystic-xxx
@alcaitiff Wow! That looks great! Good job!
No. Even the best of them is nowhere near as good as Illustrious/Pony.
@DaddyWolfgang run Flux checkpoint/XXX lora as i2i with your degenerate Illustrious smut at ~50-60%. flux composition, natural non-stiff poses, and speed all sucks, so idk why people want to gen from scratch with it so badly.