Upscaling is natively supported! Use native model+Lora loaded
Use Lanczos x1.5 (1536 res) with a strength of 0.3~0.5. The Lora is natively trained on that size!
This Lora produces full nudity:
• Breasts shapes are now more varied and work correctly
• Nipples no longer have dark dots at all, are varied, and correctly shaped about 90% of the time.
• Pussies work
• Pose concepts, crop factors & general composition was not damaged
• Does not produce watermarks
What is negatively impacted: (all of these will be solved once I add regularization data)
• Clothing (bias towards full nudity, if asking for half nudity or only slightly revealing things)
• Text generation
• Variety in faces (can be solved with prompting - but is a bit annoying for now)
• Backgrounds (yeah... way too many indoor examples in dataset. since this was just to get full anatomy running correctly)
• Men (this wont be fixed till v3, since that's when I'll add in male nudity
• Body shape (currently there is an extreme bias towards skinny)
What it doesn't do:
• Porn (flux cant do it to begin with - and I didn't train for it)
• Trans (I'm worried it might confuse the model, since the concept bleed is real. While I'll add this at sooooome point. Don't expect it till everything else is working consistently)
Compatibility:
• This wasn't trained via overfitting, so it should be compatible with every correctly trained lora.
V2: next run, will prob be on around 15k images (5k nudity. 5k normal high quality clothed. 5k regularization of completely random photography - like wildlife) using LoKr. I'll do rank 24 & rank 128 runs
----------
Do you wanna help me out?
Post a review of this model, and mention what works well for you, and what doesn't work well for you. Feedback is important!
If you wanna donate so the next model gets released sooner (basically finance my training), just contact me
----------
I'm currently working on a large flux training guide - this being one of the harder projects I'm documenting. I'll post the link when its done!
----------
This is one of my FLUX research projects. Specifically the one for anatomy training.
Trained on 4k images, at 512,768,1024,1536 resolutions, at fp16, on an A6000 for 3 days
V1: network rank 24, using standard LoRA algorithm.
Key insights:
• fp16 not worth it. will switch to INT8 training next.
• 1536 resolution was needed to remove the black dots on nipples (without relying on absurd overfitting), to add fine details on skin, and patterns on clothing. However more research is needed, since I think adding a smaller subset at this resolution will most likely be good enough. There is a very very high chance that a 2048 resolution is actually needed though. I don't want to think about it due to vram requirements.
• LoKr will be a much better training method for this. So I'm switching to that next.
• trained on simpletuner
You wanna use this on your site, or in your merge? -Contact me.
Description
can do all redheads + one custom face "regigi"
FAQ
Comments (20)
dammit. This is really good. Now I actually need to switch to flux full time.
There seems to be a tendency for random tattoos on upper arms and collarbones
Yes, that happens with normal Flux as well. Probably it's just reflecting about what's common in 2024
I noticed it as well. But its probably an issue of token bleed on my side. The "tattoo" token accidentally bled into a few other tokens due to how my captions are set up. Good news is that this is very fixable - so will be less of an issue with each update
seems like a race to generate female anuses 🤣
To the farthest reaches, where legends sleep!
a 50/50 mix with my is quite ok ;)
shame that the nudity quality is about on par with photorealistic 1.5 models. getting realistic vags would help a lot though. but its a start in the right direction.
Yeah,excatly.
I guess this are all starting try and errors because Flux is a brand new model,so people need to learn how to properly train the stuff and not to affect quality.
My guess is very soon we would get a full checkpoints that are trained on nudity,same as it was with SDXL.
it is really interesting how you have a detailed description, like you mentioned the key insights and everything, really appreciate that
Still the best of the releases so far. I keep doing A/B, A/C, A/D, A/E tests, and this one reliably comes out on top.
Love me some Ariel A
My favorite LORA. I find I can get pretty good facial variation without prompts or degradation in the results by setting strength_model and strength_clip to 0.5
This lora is a great start! Some observations I made with strength 1.
tl;dr: Makes very good nudes, but it overrides all the flux defaults and you have to like the style.
Pro:
- Breasts are firm and round and nipples upright by default (Flux vanilla's are a bit weird)
- Nipples were right 75% of the time (right size, right place, right color, but always hard - sometimes the monstrous growths of flux vanilla)
- Vags always were ok-ish (most loras don't get that right)
- very nice toned, slim body type
- Skin texture is very good already, better than flux vanilla with guidance 2.5
Contra:
- has a huge Same-Face-Problem and I personally don't like it (look at the images and you will see every second of it showing the very same girl)
- 25% of the generations have the same (!) Tatoo on the right shoulder above the breast - some additionally on their arm or elsewhere - PLEASE LOOK AT YOUR DATA
- Hands are not right very often (70%)
- flux poses are replaced with its own. It's more dynamic and amateur, less suitable for fine act
- breasts are always perfect and for one prompt mostly the same size
- skin has no flaws, no blemishes, red spots, nothing - should be more varied
- faces of full-body shots and further away need facefix.
- composition is much simpler than default flux - no blurred foreground (or displaced), no thirds-rule...
- lighting is very simple and unspecific compared to base or other loras
- Displaced or doubled arms or legs, one monstrosity like SD3 laying in the grass - but not so often.
- Always displays hard nipples - add soft ones and select with a trigger word?
- Introduces a slight grain to the foreground. But isn't that bad, this even deepens the skin texture a bit.
Meant as orientation and constructive feedback. I hope it helps someone or to make the next version even better.
hmm this unfortunately changes my character's lora face
yeah. faces have real bad bleed. Good news is that using LoKr, completely solves this. I've already tested it on small scale trainings. So once I do a v2, it will be fully compatible with all face loras & lokrs
@Caithy will be looking forward to it, send some buzz for your efforts!
@Caithy how do you use lokr?
This is great work, I get just what I need ;-)
State of "AI Art" - When the title image with "Correct anatomy" text has broken anatomy. People have no idea how actual human body should look like.