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
FAQ
Comments (37)
will there be a turbo version: like 8 step or 10 step version for fp8 version? --- it would be awesome!!!!!!
This comment is incorrect, but leaving in case someone needs it The link below only works for SDXL-based models. Not Flux-models.
https://civitai.com/models/1608870/dmd2-speed-lora-sdxl-pony-illustrious
You can use this LoRA.
I could bake it into the model, but then it would be "forced". It's better to use the LoRA on top I reckon, so people can use it either way.
6tZ will that work for Flux? as Fluxed Up [Flux NSFW Checkpoint] is flux...
ddflc27 Ah no, sorry, I didn't pay attention to which model you asked. No, this will not work on Flux. There are other ones for Flux but I never liked them, the quality was too poor.
When I try to load your workflow, the SaveImageWithMetadata node isn't downloading, and I can't find it in the search bar.
It's not really needed, just use the normal save image node :)
Not all nodes are included in the registry.
Here's the github:
https://github.com/nkchocoai/ComfyUI-SaveImageWithMetaData
Heyy, really excited for this, do you have any ideas about doing the same thing but for flux krea
Ideas: Yess
Time: Less
Thanks for this model. Looks awesome. Thanks for share the model with us.
Thank you for this masterpiece, obviously good for realistic characters, also for landscapes. Best FLUX model I've tried. Great work. Thank you for sharing it.
EDIT: My bad, we can use generated pics for a commercial project.
Just wondering, does this work with Loras?
no its overcooked
I use it with a lora trained on Flux1.dev and it's just messing with the face a bit, but if you use it for body focus is working well.
For me is a big yes, just track down some of the old forbidden civtia LORA's and test it. Try the cthulhu ones they work.
It works i promise.
Any chance of getting it quantized with Nunchaku? I've been trying to do it myself, but it seems to only work well with diffusers formatted models. I could also try it myself if you had a diffusers version of it or know how I can convert it into one.
I'm sorry but I don't know how to do this conversion unfortunately.
If you could help me out with it I'd be happy to upload it to here, or you can upload it of course :)
bump, sounds very interesting to my low vram
@yuichi001 Does the Q4 not work? I just posted an updated version of this. This should work on 8gb VRAM at least.
@6tZ The Q4 is a bit derpy it confuses some things and it has a tendency for body horror in some prompts.
I rather downloaded the FP16 and converted it to Q6, it works perfectly. From my experience Q5 M K is perfect for low vram and Nunchaku int/float 4 for ultra low vram/ram
@cesa210 Agreed. Lower Q versions are not performing as well. Once I realized the difference I sacrificed speed and never went back. FP16 is the minimum that I generate with myself, except for when sharing gguf's of course.
Possibly Nunchaku int4/float4 would be better there, but I don't know the techniques, and the day only has so many hours to experiment :)
@6tZ @cesa210 Oh yeah, I forgot I asked about this. I ended up making a Nunchaku version of it, seemed to perform decently.
@thaddeusk How did you make it?. i research a bit and the information is not very clear.
Also, i just bought a 5060 TI with full fp4 support, supposedly all RTX 5000s are, and i would love to convert some FLUX models, like this one, to the nunchaku format.
@cesa210 it was quite the journey to figure it out. Trying to get it to run inside my 32gb VRAM without taking a week to complete took a lot of trial and error. I summarized what I did in this thread https://github.com/nunchaku-tech/deepcompressor/issues/66#issuecomment-3393116934
I'm gonna be blunt and brutally honest, no time to beat around the bush.
If someone likes random results and little control over composition, this model is not too bad. But I gotta say, its not for me. If I try to get a person in a standing pose and every result is a totally different perspective I know the data set has absolutely no stable base line and no good image captioning. The image quality is actually not bad, but its not helping if the model does random stuff.
Also if one of the first images is a woman holding a cell phone in front of her face i am already half way out...^^
I appreciate the effort and am grateful people keep trying. I am mad that I don't have the resources to run a finetune with my own data, barely get the Loras done I am working on. Its really a stupendously expensive hobby. So a round of applause for the creator.
I agree with the critique. It's really frustrating to fine-tune and lose these capabilities. The problem I've had/worked on mostly is how hard it is to get NSFW into Flux.
It's certainly overfit and has extremely strong biases, to a point where prompt adhesion is hurt.
The best idea is to blend it out more with other models to not be as biased. But then you lose the artistic style that is actually good with this model.
@6tZ Its a really complex matter and its easy to break Flux. What I have figured after the past years experimenting and now the experience with Flux is, the nudity aspect in Flux is mostly a "knowledge" gap. The model simply doesn't know how to do it. That is a chance but also a tricky thing. If you train with nude images you have to fill those gaps without overfitting the rest the model already knows. There are many things one can do to address that, but all of them require extra work and a very particular targetting of the aspects that one wants to train.
I think two things sound very promising to me you could consider when planning your data set. First is abstraction, because the transformer is really good in generalising, you could increase the number of samples by adding "the thing" you want it to learn in abstract styles that you don't actually aim at, the main benefit is you can spread samples over many categories without overfitting the "photo realistic style". But then you need those photorealistic samples too, and there you can work with masked samples, or manually editing. Its very common that images tend towards certain compositions depending on the data set. And its a basic lesson we already knew with sd1.5, removing background and other noise is quite important to avoid backing in compositions. Of course variety is another thing to counter overfitting but that is such a basic thing that I take it as a given one must know.
Probably you know all that and the actual issue is that one day has only so many hours to work on a data set^^
@mad_rooky Yeah and training and working with Flux in general takes forever, and generation/testing is slow too...
Not to mention that you need so much VRAM to be able to actually finetune a checkpoint.
@6tZ Yeah, the vram, and NVIDIA is kneecapping us with that, one of the least expensive parts of a GPU.^^ But I blame AMD for not being a proper competitor... :D
What you can do though is to train Lora and merge it into a full checkpoint. It doesn't make the whole process any easier but it could be a solution to avoid burning computation directly on fine tuning. And it might allow you to balance your results better. If you need to repeat a merge it is much less of a strain than repeating the fine tuning itself.
@mad_rooky Yeah, that's what I've done on my checkpoints here. LoRA training, and then merging with it and other LoRAs and checkpoints.
For composition you could try comfyui flux depth lora, results are pretty good.
For better prompts, try https://huggingface.co/spaces/fancyfeast/joy-caption-alpha-two
Flux doesn't help. But SDXL models are getting really consistent right now, maybe give one more year.
The randomness is for me a selling point, as with most "trained or merged models" it's nearly always the same couple of faces. "Variation at the cost of cohesion is a price worth paying".
@pursuit_of_beauty I think improved prompt cohesion is always a good thing. But randomness should be encouraged on top of that.
If you prompt something, and get something completely different, the model may be overfit in a bad way, and this model is definitely that. So are most of my checkpoints to be fair. I don't really "precisely" fine-tune them. I mostly go by feeling and vibe and an aesthetic I'm looking for.
It may make them less professional, but it makes them look better.
So style over substance. Ohh, that's a good model name...
@6tZ What I've started doing to get better control over the results with this checkpoint and flux in general is to start with a pony model like cyberrealistic and generate a character and scene with the correct composition. Pony has way more loras that are easier to manipulate the scene, character and composition in addition to much better prompt adherence, though it does lack realistic faces. Then I'll run it through an image 2 image workflow in comfyui (denoise 0.4) with an upscaler about 1.5x (usually a couple iterations) using flux as a checkpoint. (You could do this as many times as you want, just keep upscaling and resampling). The results are near perfect and high resolution. This is the best method I've found (on top of precise prompting) for getting the results you want with flux. I'm sure you know all this, but just my 2 cents on what is a common complaint with flux. Thanks for the models!
@mikeeeyyyyhk7 Yeah it's a good solution.
You can also use NAI-models for the base, and use my Illustrious Realistic model for the textures, it works wonderfully and can output some bonkers stuff!
@mikeeeyyyyhk7 can't you do flux controlnet anyway