Welcome to The Fux!
The goal of this model is to unlock full NSFW capabilities in Flux, one step at a time, starting with a photorealistic female focus. Enjoy the show!
Recommended Settings
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...
Description
Version 4.2 focuses on aesthetics and NSFW with additional privately generated datasets for genitalia stability, as well as:
MysticXXX v7 by the brilliant-as-always alcaitiff
Nude Art Opus V by the astonishing Adel_AI
Aesthetic Amateur Photo V4 Beta2 by the mighty powerpuff
FAQ
Comments (23)
Hey there! Wanted to give some feedback. 4.0 is amazing. Lora work perfectly, in most cases better than with dev. Body horror happens if you prompt in a way it does not like, but prompting has a huge effect and can fix it most cases. 4.1 lost something. My Lora look less sharp, likeness tanks and the prompt adherence is quite bad. It does however have less body horror, but prompting seems less effective over all. It can produce good results without the lora, but for me lora are vital for characters. Could you provide a changelog in this or future versions? I would love to know what your intended use case is for each model and how they all differ from each other. Thank you over all for your hard work, it truly shows in your models. 4.0 is quite possibly my favorite checkpoint of any base model so far. <3
Thanks for the feedback. It's always appreciated!
I find that I don't have time to share the process and details enough in general, but I will help when I have the information to share (if I remember).
Version 4.1 focuses on improving faces. It merges in various LoRAs and some custom training to improve faces. Less Fluxface, and more face variety.
The SameFace Fix by VadimFedenko is included at -0.35 weight.
Version 4.2 focuses on aesthetics and NSFW with additional privately generated datasets for genitalia stability, as well as:
MysticXXX v7 by the brilliant-as-always alcaitiff
Nude Art Opus V by the astonishing Adel_AI
Aesthetic Amateur Photo V4 Beta2 by the mighty powerpuff
Regarding the worsened LoRA quality, I'm not sure. I feel like it should not have been affected by this. I don't test all versions for all cases, so sometimes things like that happen.
I'll make sure it works well in upcoming versions by blending it out a little bit with the original model, or earlier versions where it worked. I've had this problem before and managed to fix it.
If anyone knows more about what causes this problem, do let me know.
I hope this helps.
@6tZ I think it is the sameface lora having some effect on the faces! Not sure though, but I do notice when I use that lora on dev it muddys the faces. It does work as intended, with some downside in fuzzy edges and a specific style that bleeds into the other lora. Thank you so much for your response!! I am excited to see a 4.2 posted!! Thanks for making and sharing these so quick! A changelog does suck to write, but it really does help someone to decide what to use and when. Thanks again!
@fromshahnahnahb957 Agreed! Just a bit of limit to the hours of the day :)
Happy to share of course.
I mostly focus on the output visuals. I mostly train by concept, and merge by feeling.
@6tZ I like having multiple versions 4.0, 4.1, 5.0 etc and picking the one that fits the job, each one has different styles
@sexgod1979 Same here! Though I don't know what they are all good for, I just go with vibes and whatever feels good at the moment. But I spend most of my time testing new versions of course.
@6tZ its working well, keep up the good work
Good model, but it still has the static flux breasts look.
Static flux breast look, not quite sure what you mean, can you give some examples with/without?
@6tZ They all look the same, no variance. No Torpedo, no perky etc.
fux is my new favorite toy. keep up the hard work
Why is the fp16 pruned 4.2 UNET model over 22GB, when the basic flux1D pruned fp32 UNET is only 16.6GB? And not just for this fine-tuned model but for all the others.
No idea. Math?
Pruned FP32 sounds like it's not full precision?
@6tZ He does have a point: are you sure you trained based off of the pruned checkpoint instead of the full one?
The full BF16 one on BFL's hugging face is 23.8 GB, so it does sound a bit like you might have trained the full one instead (https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors)
Edit: (This actually works out even better for me personally since I use the full one, but I just wanted to point out what it appears happened)
@ThrowawayAccoun42414 I honestly dunno, all these numbers are giving me headaches...
@6tZ Not much to do about it for now regardless, unless you want re-tune it. Maybe just something to check out for the next version? 🤷
@ThrowawayAccoun42414 Yeah. I haven't really trained on the core model. I train LoRAs and merge them in. But at this point I'm months into the model and I can't remember the source ^^
@6tZ Oof, fair enough! You may want to consider not labelling it as "pruned" next time, but it's up to you!
@ThrowawayAccoun42414 I'm probably not using that option/term correctly then! Sorry about that.
These things are so technical and confusing. As a rookie it's hard to know what's what.
Any guidance is appreciated. Why should I not use pruned? When should I use pruned?
@6tZ well, in this case I'd say to use pruned when the fine-tuning is based on the pruned flux model :P
A pruned version of a model is sort of (in simplified terms) exactly what it sounds like: some less important/low-impact parts of the model are "pruned" out and zero-ed out/removed. This lets the file compress to a smaller size and let's the model run faster by skipping those parts.
@ThrowawayAccoun42414 I thought any low precision formats counted as pruned?
i.e. fp8 is pruned because it's not fp32?
@6tZ No, that's quantized/quantization. That's different than pruning
@ThrowawayAccoun42414 Gotcha, okay.
So a model would literally be titled "pruned" for it to count as pruned then. And doing this removes some detail, for some other benefit (performance/size I assume).