See my examples for prompting style. Just use natural language.
No specific keyword but words like nipples, breasts, and topless work.
The samples were generated with LoRA strength 0.8.
I find it helps to give random names, especially if you want multiple people.
I've been sampling with guidance 3.0, simple euler schedule with 40 steps. I was using with max_shift and base_shift 0 but other values work well.
It is trained with the full precision t5/clip conditioning so it will work best with those.
Feedback welcome.
Description
It was trained with 500 images on an H100 with batch size 4 for 20,000 steps (160 epochs). I used quantized model weights for this version but i'm testing another one without. Rank 128 and alpha 128.
FAQ
Comments (54)
Great progress!
Rank is way too high.
why do you say that? just cause it's a large file? i find larger rank == better model, up to about 128
Maybe it has to be. All the small lora aren't working as well.
@ryanjay0 I say it because the rank is too high. For this size it could be a general nudity lora, not just for topless and breasts.
@Gore_Man The small loras by punzel are working as well as the 150/300 MB loras by others.
@mrneon Which one is that. So far the best for me has been the "flux improved female nudity". Don't want lora to overpower faces, lose styles or cause body horrors.
@Gore_Man i haven't seen anything from that model that approaches the realism of this LoRA, but maybe i just like this style better.
@Gore_Man The celeb loras that punzel has been making.
@ryanjay0 They are very different loras, your dataset is clearly superior with a bigger variety.
Using flux1-dev, I'm getting pepperoni (super bad) nipples.
is that bad: do you mean weird nipples or just large areolas? what is your prompt? are you using the 16-bit t5?
try some of the prompts in my sample images with those settings
Yeah, sorry, not a compliment. I mean I'm getting either the weird button nipples or the completely flat pasty things.
I've grabbed your first image and thrown it into ComfyUI.
I needed to fix and replace a couple things (different LORA name, Flux1-dev name and VAE name). By default, the "Set Parameters" section LoraLoaderModelOnly has no model input. I assume that should be connected to the output of the Load Diffusion Model, so that's what I did.
Also, I'm not clear on why you've basically got a duplicate flow down below. What's the reasoning for that?
Since there's no seed in posted Comfy pics (from what I've seen) I'll run a few and see, but
@meritrash6350 that's just a flow to compare loras.
@ryanjay0 Ah, didn't realize. OK, that makes sense. Anyway, I posted one sample, and I could post a few more, but they're all the same - tiny button nipples. I also posted a shot of the flow in case you had any suggestions.
@meritrash6350 some of my newer uploads have seeds. id start there and see if you can reproduce what i get.
@ryanjay0 OK, this prompt: Carrie and Eileen at Mardi Gras topless with lots of beads around their necks. Carrie is black and Eileen is a redhead
Seed: 182412226563642
I've bypassed everything on the lower flow since it's not like I have ToplessV2 anyway. I'll generate a couple and see what I get.
For some reason (and I'm starting to wonder if this is indicative of whatever is going on here) whenever I grab one of your flows from this, or someone else who has posted to this LORA, the input for the LoraLoaderModelOnly is ALWAYS connected to nothing. I assume it should be connected to the output of the LoadDiffusionModel....but it sure seems suspicious.
I was getting these kinds of issues in the CLI: transformer.transformer_blocks.18.norm1_context.linear.lora_A.weight
but an update seems to have fixed those
Holy shit, that is impressive as hell
Thanks.
Spectactular. Nothing else is coming close to the quality of nudes I'm seeing from this LoRA. The realism in particular is truly mind-blowing. I'll have to post some pictures. Whatever you're doing, please don't stop making Flux nude LoRAs!
I'd love to know a bit more about your training process. What scripts did you use? What guidance scale? How many photos? What did you use to caption? Did you use regularization images? Maybe asking for a write-up is premature but I'd love to read it if you were to make one!
accidentally the best part of this is the sharper backgrounds.
Amazing model - and we're only a week into flux. It really nails realistic details better than SDXL and this is just a LORA!
I think there's room for improvement in size variation. There's not a great way to make smaller chested women - likely just due to training data.
Great work!
the other nsfw loras don't compare. this is the GOAT
thank you.
by using your exact workflow for this image i get totally different results :(
nearly exact same as this: https://civitai.com/images/24288932
i edited the metadata to add the seed. that should help... although i get similar ones with other seeds. i just posted another one with a seed also and a simpler workflow.. good luck. i'm also using full-precision everything. so if you're using anything quantized that might explain the difference. Also make sure to update comfyUI. it recently added support for these LoRAs
Man... you deserve a medal... the comfyui update made the thing! THX!!!!
This is a fun addition!
Works awesome.
With Forge and dev_Q models the LORA introduces some kind of noise into the image. It can be seen on some of the example pics as well.
Thanks for the model, it works great. If you could add more skin details, it would be even better though. The body overall still feels a bit plasticky.
I'm curious what your mix of image style types for the 500 source images were...like what percentage were photos vs. others types?
all photos.
Would love to use it, but it's too big. I don't have this much room RAM wise.
unfortunately i won't be making smaller ones since i think it impacts quality. you can always pre-merge the LoRA weights and then it won't take any more ram than your base model.
@ryanjay0 can i ask how to pre-merge the lora into base model, thx
@BNP222222
you can do it with some code that looks like this using transformers.pipelines library in python:
pipe = FluxPipeline.from_pretrained(model_name, torch_dtype=torch.bfloat16).to('cuda')
#pipe.enable_model_cpu_offload() #save some VRAM by offloading the model to CPU. Remove this if you have enough GPU power
pipe.load_lora_weights(lora_folder, weight_name=lora_name)
pipe.fuse_lora()
pipe.unload_lora_weights()
pipe.save_pretrained('merged_model')
God approves this Lora. God loves boobs. He created them.
Finally! More bobs and vegana!
This is still one of my favorite NSFW flux loras. I published a workflow using it to make realistic amateur photos here!
where I can use this nice nsfw flux? Have U recommended best premium web AI NSFW generator with refiner and complete tools?
thanks for saying it's nice. i use locally on my gpu with comfyUI but you can generate on civit also. i dont know about something with refiner and complete tools. you can basically do everything localy if you are willing to learn and have a decent gpu.
Unfortunately I have not good PC to run :(.
Thanks for answer my comment ^.^
@ryanjay0 I've tested a few dozen LoRAs now for Flux, and this is one of the very best NSFW. It is a very large LoRA at 1.3GB, but whatever you did to train it, it worked very well. Maybe it is so good because it is so big in filesize, capturing many details and diversity that base Flux lacks? Can you share more of your process? I look forward to version 2, and other models you develop!
thanks so much for saying that. i do think larger loRA size is better. this is rank 128 with 128 alpha. rank 64 also seems good but i think it gets significantly worse as you get lower than that. i also trained a lot of epochs and mostly used wd14 comma style tags. i also prefer an older version of ostris trainer as i can't get the same result now.. boring reality creator said something similar.
This worked great! I love the variety of different sizes and shapes
Breasts have a great variet of shapes and sizes, but it always generates Flux standard giant areoles.
i dunno if that's flux standard, but i'll admit that's what my model does. thanks for the feedback. maybe someday ill make an update that listens to more breast related attributes.
Details
Files
flux_topless_v1.safetensors
Mirrors
flux_topless_v1.safetensors
flux_topless_v1.safetensors
flux_topless_v1.safetensors
flux_topless_v1.safetensors
flux_topless_v1.safetensors
flux_topless_v1.safetensors
flux_tl_v1.safetensors
ft.safetensors
flux_topless_v1.safetensors
flux_topless_v1.safetensors
flux_topless_v1 (1).safetensors
flux_topless_v1.safetensors
Flux_Topless.safetensors
flux_topless_v1.safetensors
flux1dev_topless_v1.safetensors
flux_topless_v1.safetensors
flux_topless_v1.safetensors
flux_topless_v1.safetensors
flux_topless_v1 (1).safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.