Welcome to The Fux!
The goal of this model is to unlock full NSFW capabilities in Flux, one step at a time, starting with a photorealistic female focus. Enjoy the show!
Recommended Settings
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...
Description
FAQ
Comments (45)
5000 buzz? OK but should be 2500 if it's only half working. :-). mostly joking but kinda not joking
You're not wrong!
I'll probably lower it in a few days, but for now, let the needy sponsor the future :D
What the flux is going on with the pricing, even be it as 'virtual' as 'buzz' ? CivitAI, aren't you afraid of legal retribution if someone reports exchanging dev models for any kind of currency? This should stop, now!
This place used to be open and free...
@darebgd388 This looks like such as awesome effort, but I too worry that the Non-Commercial clause would prohibit this.
@darebgd388 openness and freedom has never lasted. Anywhere. Ever. Stay vigilant.
It's awesome that the bf16 version is available. But right now
ComfyUI TensorRT can't be converted successfully yet. Hopefully it will run efficiently on 24G Vram by the time the model early access is over.
@darebgd388 It is free. Also, technically, nobody is paying money. They are using buzz. Also, be patient if you want free. Quit bitching with your dick in your hand. Go make some money, erm, buzz.
@Gradasho it is not free. buzz costs money. there is no way someone is earning 5k buzz the free way. it would take several weeks of spamming reactions/pictures to even get close to 5k
@darebgd388 @mikemichaelguillermo CivitAI has a commercial licence with the creators of Flux that covers the site, uses, and buzz surrounding finetunes of the model.
@suede2031691 It runs fine on 24gb vram. BF16 is the way the original model was released, it's what I used to finetune.
@6tZ That's very cool and I stand corrected. I did not realize that CivitAI had an agreement with Black Forest Labs. Thanks for the model!
Doesn't work very well. I keep spending points on these things and getting very little bonus from the last iteration. I'm done buying them.
it's too early for actually good trained models. give it a month and stop spending points.
Sorry to hear that, but thanks for the input.
Can I ask what it is that isn't working very well for you?
I assume you're getting it up and running in the first place.
It is indeed a work in progress and I've shared images with the results you can get from it at about a 50% accuracy.
@6tZ It has little to no pose control overall. It has very tenuous grasp of concept, loses a lot of information on scenes, and overall does a pretty fixated job on single ideas.
Flux HAS all of these details built in, they just need to be properly accessed. This lora seems to be hindering that access rather than helping currently. Retagging using pose and angle control would help a lot.
@AbstractPhila Thanks for the advice. That's a very good observation. You are absolutely correct in it.
I'll be doing several more re-captions to experiment more on it, and I will likely train versions that mixes it.
Right now it's all long and complex prompts that focus equally on the character's description, the character's action, and the background.
I think this is the wrong way of approaching teaching the models new concepts.
It can get lewd images in the poses that are trained into it, but it doesn't add understanding so it's hard to work with.
I appreciate you spending the buzz, thanks to you and many others who were curious to test it I now have enough to train 50 more LoRAs with experiments and hopefully contribute with a stronger model in the future.
I hope you don't feel mislead by the model. I feel like I was perhaps too clickbaity in the title, but clear in the model description.
Regardless, thanks for your feedback, it really helps!
@6tZ This character poser I'm currently generating images for should do the trick to pose anything. So stay tuned. I won't charge buzz for it.
extract a lora from the sdxl anteros model (nsfw sdxl model) then use a saveclip node on your extracted model to get a clipl model out of it from comfyui, then put the model in this modified node and turn the clip_l strength up to 100. Makes a big diff for me. https://github.com/311-code/ComfyUI-Flux-clip-strength
@user1234123 I've experimented with Everclear, Autism, and Realism CLIP_L's from the pdxl crew. They usually work at a lower power as long as I have my PDXL consistency v1.1 loha merged into it. They often introduce nsfw and style elements rather than context though, as the SDXL CLIP_G is where the majority of the context is.
Bad pussy, bad anus......sd1.5 is better... not you fault, its a FLUX thing.
@tupu I have discovered a training method, that does not make it a flux problem anymore. I will be sharing my findings soon.
Update: NVM. Thought this was an all-in-one.
ComfyUI's "LoadCheckpoint" Node does not recognize either model as a valid safetensors file. Share workflow?
Good point! I'll add a workflow.
Please try with the "Basic Workflow" from here:
https://civitai.com/models/618997?modelVersionId=691988
I'll look into redoing the model in formats that work more out of the box. I was still using this format myself so I haven't messed around with the others yet as it's moving so fast.
Sorry for any trouble, send me a PM if you still can't get it working!
Yeah, pretty disappointed after dropping 5k that this wont import the same as the other Flux models I've downloaded. Thinking maybe I need to try the other version you've uploaded but I'll wait.
Hmm, no the other one would be the same. It's just saved with different precision.
Are you using ComfyUI?
Please try with the "Basic Workflow" from here:
https://civitai.com/models/618997?modelVersionId=691988
I'll look into redoing the model in formats that work more out of the box. I was still using this format myself so I haven't messed around with the others yet as it's moving so fast.
Sorry for any trouble, send me a PM if you still can't get it working!
>Dropped 5k buzz on a model that should have been free since first day of release
Yup, you're part of the problem too.
If it's 5000, pony is also a good choice
@AIexperienceguy > "Yup, you're part of the problem too." L O L.
I can't speak for the asset here, but my content I've uploaded is heavily trained on my own personal datasets of photography I've shot over the course of my 20 year professional career. $5 is absolutely not an unreasonable price for assets that have had a lot of work put into them, or content that is trained on new unique datasets. Sure, there are plenty of lazy ass 50/50 merges on here that don't deserve anything, but a lot of this content, especially stuff that is trained on new datasets, COSTS A LOT OF MONEY AND TIME TO MAKE.
But your opinion is real rich coming from someone that has ZERO UPLOADS and hasn't contributed anything at all to the community he expects to get everything for free from lol. It's ALWAYS the moocher accounts like "AIexperienceguy" with these hot takes. What a joke.
@AIexperienceguy Translation: "I feel entitled to get everything for free, how dare you pay for it and be part of the problem of rewarding those who work hard for their work??"
Let's get started with porn! In 2 weeks when it's not 5k to try an experimental WIP model lol
Fair :)
@6tZ :)
I'm more than happy to support creators. That being said, asking for 5k is ridiculous. I would be more than willing to pay 1.5k tops for early access.
@KelevraQuakenstein Totally fair and thanks for your input.
I don't quite know where to put the price point! Maybe I'll try a cheaper version for the next version, or a different sidestep model I have coming up.
In the end 5k is $5, which so far ~30 or so people agreed to pay for the model. I understand this is because it's Flux and there's no "solved" NSFW yet, so of course I'm taking advantage of the market in some sense. Capitalism?
But it is also to compensate for the hours and days of work and training time that goes into making models. It's based on a dataset created specifically for this model of 24k images, with custom scripts written to manage the dataset with various styles of captions, to figure out how to train NSFW and concepts into Flux, which is not really taking it on like a champ. Some payout of the hours of frustrations and 30-40+ failed model trainings and painful testing, to finally make some progress etc.
And lastly, it has to pay for my psychologist visits, with the amount of body horrors I have endured for the fine folks at CivitAI. I've seen limbs coming out of EVERY part of the body, and being bent at exactly all the different angles that are possible (and more like, impossible). When you get access to the model, try prompting for "penis" and you will have a blast :P
needs text encoders to work on Forge, although not sure if there are ones other than recommended by Forge?
I've tried one or two prompts from the images above, they're a no go.
Are you getting completely different results? Is the model getting loaded properly in Forge?
@6tZ yeah, a few looked completely different or were pretty much censored i think/ don't know if it's any other settings I need to change/add. As an example I tried the Sierra Leonean woman and it comes up with a pink hue.
I've done one with my own shorter prompt which comes out ok but think it's not quite correct (uploaded to the gallery)
also uploaded the below.
The image is a high-resolution professional (film grain:1.2) photography of a 30y-o Sierra Leonean woman with medium breast size, Black drill hair hair and a screaming orgasm expression. She has very realistic skin details, extreme realism. This image is a highly detailed, digitally rendered illustration in an anime style. It depicts a young woman with long, dark purple hair and fair skin, lying on her back on a tiled floor. She has a slender physique with medium-sized, firm breasts. Her eyes are closed, and her mouth is open, suggesting an expression of pleasure or ecstasy. She is wearing a provocative outfit consisting of a short, tight blue skirt with gold trim and a red bow tie, paired with a white blouse with gold trim and a red bow. Her outfit is partially undone, revealing her bare chest and a small, red gem on her skirt. Surrounding her are numerous large, pink tentacles, some of which are inserted into her mouth and others wrapped around her body, including her breasts and genitals. The tentacles are detailed with a glossy, slightly wet texture, emphasizing their phallic nature. The tiles beneath her are clean and white, adding to the stark contrast with the dark tentacles and her pale skin. The overall composition is intense and erotic, with a focus on the interplay between the woman and the tentacles. The image is highly polished and detailed, with a high level of realism in the textures and shading. intricate details, warm dim lighting, nipples and pussy, dynamic angle, (film grain, high contrast, sharp details)
Steps: 20, Sampler: DPM++ 2M, Schedule type: Karras, CFG scale: 1, Distilled CFG Scale: 3.5, Seed: 956437839704082, Size: 896x1152, Model hash: c0da935cf8, Model: fuxCapacityNSFWFluxBaseModel_v03Fp8, Version: f2.0.1v1.10.1-previous-418-g0f3309eb, Module 1: t5xxl_fp8_e4m3fn, Module 2: clip_l
Flux will take a lot longer to get to pony levels. I think in about a year we'll see something that looks good.
Skin looks fake, like everyone's a blowup doll. I mean, if that's what you're in to, well...
try lowering the guidance and denoise
Looks like Pony... I bet they used it for training images.
@diogod Indeed, it's a mix of SDXL + Pony for dataset.
A solid effort, keep up the "hard" work!
mmm so hard.