Now also available as a LoRA, with more images and training tweaks: https://civarchive.com/models/1031648?modelVersionId=1157036
NaturalBeauty FLUX v0.1 is a photorealistic FLUX.1 [dev] base model for creating full-body nude and topless images of beautiful women. NaturalBeauty enables FLUX to reliably generate naked and semi-naked photos in four common full-body glamor poses.
To encourage nudity, add
nakedortoplessto your prompt.To unlock the four body poses, add one of:
standing, wide stancestanding, feet togetherkneelingsquatting
This initial version of NaturalBeauty for FLUX is the result of a year-long project to tag nearly 11k images from a database of 1m candidates, using a diverse and flexible dataset with full tagging of background imagery. No merges; just images and hand-tuned training.
Tagging
NaturalBeauty FLUX performs best with a prompt that follows this format:
{a caucasian|a black|an asian|a latina} woman,
{standing, wide stance|standing, feet together|kneeling|squatting},
{naked|topless},
{barefoot|high heels},
{short|medium length|long|very long}
{braided|braids|crimped|curly|ponytail|straight|wavy}
{black|dark brown|medium brown|light brown|brown with blond highlights|chestnut brown|auburn|colored red|bright red|orange red|orange|red|strawberry blond|light blond|medium blond|bleached blond|platinum blond} hair,
{neutral expression|slight smile|full smile},
mouth {closed|slightly open|open},
eyes {closed|open},
looking at camera,
{tiny|small|medium|large|huge|massive} {fake|natural} breasts,
{arms by side|arms folded in front of body|hands behind head|hands by shoulders|hands holding breasts|hands on thighs|hands on hips}For example:
a caucasian woman, standing, wide stance, long wavy medium blond hair, slight smile, mouth closed, eyes open, looking at camera, small breasts, hands behind headNaturalBeauty FLUX is optimized to produce good results without needing a negative prompt.
Recommended generation settings
NaturalBeauty FLUX is trained to generate full-body portrait images with a size of 832x1216 pixels. Other 3:2 aspect ratio sides should also work well.
Training settings
10,937 hand-tagged images
batch size 20
10 epochs
constant learning rate of 1e-5
1x H100
kohya-ss scripts
no merging :)
Description
Initial version with support for four common poses.
FAQ
Comments (38)
Please provide quants and fp8
I'll read up on how to do that, and hopefully provide them soon!
I believe there are a GGUF conversion scripts available on hugging face, but don't remember precise link or keywords for search.
GGUF Q6.* or Q5.* are more profitable quants in comparison with FP8.
I've uploaded an FP8 conversion, and I'm seeing results in my own testing that are almost identical to the BF16 original 👍
I'll keep investigating GGUF, but in the meantime, please do try FP8!
@BelieveDiffusion GGUF Q6.* usually has a performance the tiny bit lower than FP8, but has significantly lower size. So, if you satisfied with FP8, you should be satisfied with GGUF.
@homoludens Thanks for the tip and extra info! I will check it out.
@BelieveDiffusion But please don't remove FP8 if you agree to let the users to make a merges. As I know we still don't have a scripts to make a merges or perform a fine-tuning with GGUF. Only FP16 / FP8 training and merging is possible so far. (if my knowledge not outdated)
can this be distilled in a LoRA?
I'm going to experiment with that next 👍
@BelieveDiffusion amazing!
LoRA version (with more images and training tweaks) is now available: https://civitai.com/models/1031648?modelVersionId=1157036
for some nude 22GB model, WHY ???
he can't say anything positive unless it's his own model. He literally just shits on everyone. I refused to use anything he makes because of his attitude.
for some 'fake' looking nude 22GB model WHY !? , in same time older initial lower requirement models can do more realistic pictures ..
@amazingbeauty One obvious reason is so other tuners can merge. You don't like, don't download. Simple. Don't be a jerk to the people who work hard to make things.
Dude. When someone spends the time to make a thing, why not give advice or constructive suggestions for improvement, rather than just dissing their work?
@funidea yeah that's not what sevenof9 did or does. He shits on any model that isn't his. He's as narcissistic as Trump. I've seen people request a fp8 nicely. That's not what he's saying. You can tell from the context of his poss to other models over the last 2 years. He's not a nice person.
its a question ... its not shit ... and if there are opinions about my models ... i'll answer them too
and btw not this model, but for flux there are so many stupid mergers, and a lot of internet traffic, that we all have to pay for this platform in 3 years... if that's what you want ???
ho hoo hold on guys this world is fake no one listen the right way to any discussion .. no wonder
we said the result is shit , people understand them as 'negative' shit ..no wonder again our whole world is fake
@amazingbeauty I'm upset about my country's election. I can't believe how many people voted for that turd. Everything about him starting with his hair and skin is fake. Everything he says is a lie.
@sevenof9247 All you have to do is write - hey can I get a fp8 version? But from your other comments elsewhere I don't think you really want it. It's not just here, but every comment of yours is on the negative. I still prefer to have the fp16 file for further merging and best details. This is perfect for me.
@sevenof9247 I just uploaded an FP8 conversion, which I really should have done from the off. The results are almost indistinguishable from the larger BF16 version, and it's "only" 11GB 😛
I'm also trying some LoRA extractions from the checkpoint. This 10k image dataset is kinda general-purpose, and so far with FLUX training, I've found that fine-tune results on this dataset are better than LoRA results. (That's why I posted a fine-tuned checkpoint originally.) But if I can get good enough LoRA results for a reasonable dimension size, I'll post a LoRA version too.
@BelieveDiffusion thx, i see, fine-tune a model I have never done before ... you use kohya?
@sevenof9247 Yep, I use the sd3 branch of kohya-ss/sd-scripts, which has his latest FLUX and SD3/3.5 additions. I used the FLUX fine-tuning script from there, running on Runpod, to train this checkpoint directly on top of FLUX.1 [Dev].
@sevenof9247 Now available as a (large but good) LoRA, with more images and training tweaks: https://civitai.com/models/1031648?modelVersionId=1157036
Thanks for the release. It says 5,470 steps with 10 epochs. The description says 10,000 images. Is that step count off?
Ah! Each step was a batch of 20 images. The exact stats are:
- 10,937 images
- batch size 20
- 10 epochs
- constant learning rate of 1e-5
Do you remember how long it took to run on the H100?
@uprz23736 Just over 26 hours. Next time I hope to try running it for more epochs with a lower learning rate, and / or a non-constant scheduler, as a comparison.
no pubic hair,no nature. by the way ,this should be a 200mb lora can handle it,why 22gb? hontsly,I had to say,rdcls.
I started with a fine-tuned-checkpoint, but I'm getting the sense that that isn't what people want for FLUX 😅 I'll investigate a LoRA too.
An FP8 version of the checkpoint is up, which halves the size for virtually no visual difference.
For this general-purpose dataset, I'm seeing better results with fine-tuning rather than LoRA, but I'll see if I can tweak things to get a decent LoRA from it too.
LoRA version (with more images and training tweaks) is now available. It's large, but that's because I prioritized quality over size. https://civitai.com/models/1031648?modelVersionId=1157036
@BelieveDiffusion great job!
not one of your models available for onsite generation? Sickkkk
I just released a LoRA of this dataset (plus more training), which is available for generating on Civitai. https://civitai.com/models/1031648?modelVersionId=1157036
@BelieveDiffusion your a real peach, thankyou for your service
@BelieveDiffusion just getting to use that one you recommended, bravo! and if my original comment came across like i was being genuine, i apologize and in fact am not hyped at the limited amount of your work currently approved/sanctioned by Civtai for onsite generation. keep up the killer loras