Welcome to Fluxtastic! 🎉
I'm excited to introduce my very first Flux DEV LoRA to you all—it's also the first LoRA I've ever published.
Now FLUXTASTIC v3:
In this version, I used new training parameters, removed all remaining AI-generated images from the training dataset, and added entirely new high-quality (real) training images while discarding the previous ones that were not up to standard. Feel free to test it out! :) Maybe share me some Images on x.com also!
Now FLUXTASTIC v2:
In this version, problematic AI images were filtered out, and the model was retrained with a different learning rate and a large set of high-quality, real images. This should provide a strong foundation for quickly creating effective and high-quality NSFW images with FLUX.
------------------------------------------------
Give it a try! There may still be some errors, but it's a promising step forward in achieving better NSFW results with FLUX.
I trained it on over 400 images, which I meticulously reworked by hand to ensure they were perfect for training.
My main goal was to remove the censorship in Flux.
For the sample images, I used a straightforward ComfyUI workflow with a Flux Guidance of 4. The LoRA strength is set to 1.
I hope you enjoy trying it out! Let me know what you think. If it's worth it, I'll continue refining the LoRA. My initial tests went really well. 🙂
Description
V2:
Removed problematic AI images (problems with the hands etc.)
Re-trained with a different LR
V1
Initial Release
Based only on selfmade AI-Images
FAQ
Comments (25)
why using ai images and not real images?
Because I used it to create my own complete data set in order to uncensor as quickly and effectively as possible.
Real images are owned by real people.
@Nitrowing except when real people purposely upload them royalty-free - which isn't as convenient as regular nudies but do still exist in vast numbers
Good Job !
Hey! Thanks for your Lora! Works fine. I'm using it a bit under 1.0 strength to be a bit more free with the type of people I can create. But I'm not a fan of using AI generated images as training Data. Your images may be very good, but will still have slight mistakes. Flux will add its own mistakes ontop of already slighty wanky images. So as a cold start it works fine to make an experimental first step, But if you plan to continue, then please switch to a set of real photos. Im mean, naked people should be one of the easiest thing to find on the internet ;)
I saw a model that was improved by realistic 3d renderings ontop of real photos because the dataset to find online was super limited for that concept. But naked people is certainly not a limited dataset ;) Just think about the future when someone trains his model on Flux generated images, which was itself trainied on AI generated images. You will introduce mistakes that are themself based of mistakes until it becomes unusable. So in short: Thanks for your first shot that certainly works as a first step, but please switch to real images for the future ;)
Hello!
Thanks for the feedback! :)
I myself wanted to test what is possible from scratch first.
I have now fed the data set for V2 with additional high quality real images and removed problematic AI images.
Currently also training with a different LR and more image material and steps. I will upload ASAP.
Will be fine-tuned bit by bit.
it's not "censorship" if the model was never trained on it 🤦♂️
Yeah, I hate how the word "censorship" has lost all meaning in the community.
You are right. Black Forest Labs simply didn't train genitalia from the start. Does that make it better? :)
Ofcause it is censorship if the reason not to train it is was a choice and not a coincidence.
what's the difference? If they chose not to train it on certain images, it is a conscious choice motivated by ideological bias.
The model was pretty clearly trained on images of topless women. Using the base model it's not hard to get such images. However the nipples are horrific/freakish. This tells me that there was some kind of active censorship going on. My theory is that they combed through the training set with a segmentation model to blur and distort any nipples.
This is an excellent lora and the model has learned a lot about nudity, both top and bottom. Unfortunately it has forgotten a lot of other stuff in the process. The struggle is real.
I'm already working on V2.
The problematic images have been removed which in my opinion were responsible for the failures.
I have continued to feed the dataset, also with real images :) will take about 2 hours then I will test and upload.
@PromptoAI Awesome. I'll keep testing it.
edit: it's a lot better.
@PromptoAI nice good to know about the dataset. In your experience is the caption or image more important in terms of dataset quality?
Flux is still fighting back. Any tips for getting rid of genitals out of the frame, crossed legs or lying on the stomach (even with "lying on back" in the positive and "lying on stomach" in negative)?
I have also noticed this problem. Try a more precise description such as "...the person can be seen in the focus of the picture and has one leg stretched out to the left and one leg to the right, the arms are leaning backwards in a relaxed position, no items of clothing can be seen in the photo"
@PromptoAI I'm afraid that even if it's highly impressiv that flux can follow these prompts, it will break the token limit really fast... :(
Does anyone else notice, using using LoRa tend to give malformed anatomy again? I put the yolov8xworld2v.pt detailer in my flow again that tackels it. But I am curious if others have the same problem.
You Downloaded the V2?
Was trained with real HQ images. What poses are you trying to create? Spread legs only works for me 1 out of 10 😬
Poses dont really matter. Its not only with this lora
Yup, lora's are breaking anatomy. I've made a handful of flux lora's so far, and the stronger the lora effect, the more difficult it becomes to produce good hands and feet, whether with my own, or any other lora. flux trains very fast, so I suspect there is a much narrower window for convergence. this will get ironed out as the training process is refined.
@gurilagardnr Ah I see. So I am not crazy. Thanks for letting me in on the dirty secret. ;-)