<h1 id="'sloppy-toppy'-concept-flux-dev1-lora!-sy4lhw5vx">'Sloppy Toppy' Concept Flux-dev1 LoRA!</h1><hr />This is a LoRA to recreate a super messy and nasty blowjob. If you're like "wait what does that mean", I'd first say "Hi, welcome to the internet. it's pretty weird here sometimes, so buckle the fuck up". After that, and if you really are somehow over 18 and don't know what that is, I'd explain: "Those hardcore facefucking blowjobs that make you uncomfortable to see the first time you see one because of how rough and insane they can get that have stuff like the man holding the woman's head all the way down til her face is flush against his pelvis, the kind where she stretches her own cheeks open while the man tries to stuff it as far in her mouth as she can take it, or the ones with tones of saliva, spit, drool, and sometimes even snot, that winds up smeared all over the man's balls and woman's face and results in her makeup being ruined and smeared and streaked from the tears from gagging and choking on the dick. Sometimes she has bloodshot eyes". This can do all of those things. You've been warned. Simply add "slptp" to your prompt. Or "slptp style blowjob", and that's all there is to it. This was trained on Flux-dev1, but it works on all the hybrid checkpoints I've tried, as well as the AIO (all-in-one) Schnell models that have the VAE, clip-l and t5 (text-to-text transfer transformers) text encoder baked into the model to make those gigantic 20-some-odd GB models that are only N4, but still the same file size as the original -dev1 weights. Of course it works best with flux-dev1, as that's what it was trained on, The dataset was tagged by hand, I used 40 images, and used florence to caption them, then went through by hand and beefed up/added/fixed any of the inconsistencies. Every caption started with "slptp", so just add that anywhere and it should give good results. I also described them all as being a "sloppy and messy deepthroat blowjob". <h2 id="to-prompt-exactly-how-i-captioned-you-should:-iegp3s1d8">To prompt exactly how I captioned, you should:</h2><ol><li>Use the trigger word,"slptp", as the very first thing in your prompt, followed by a comma</li><li> An NLP (Natural Language Prompting) prompt where you describe exactly what you want using the same language you'd use if you were writing and describing the picture for a book for other people to read, except do not use punctuation until the end of the prompt, where you'll add another comma.</li><li>Add your tags, separated by commas (e.g. "oral sex, fellatio, blowjob, nasty, messy, spit, saliva, deepthroat blowjob, a man getting a blowjob, a woman giving a blowjob, sloppy and messy deepthroat blowjob, blonde, blue eyes, perky tits, thicc, etc.... etc...)</li></ol> That should give you fairly close to the images I was able to generate.This is my first NSFW / "real" model I've added here. I've trained LoRAs forever on Kohya, but when flux came out, I edited the <a target="_blank" rel="ugc" href="http://run.py">run.py</a> script from AI-Toolkit and adjusted the config.yaml to meet my own needs. I actually made a google colab book for it, which you can find here. It's not much, and all credit absolutely goes to the creators of AI-Toolkit, as all I did was experiment with settings until I found what works for me, which I've left as the default options. If you want to thank anyone, thank them. Okay! Happy gross/nasty/sloppy/messy/rough blowjob generating! <h2 id="how-to-install-comfyui-and-use-this-lora-with-windows-11-and-nvidia-cards-only:-447sfzvjq">How to Install ComfyUI and use this LoRA with Windows 11 and NVidia cards only:</h2>(You can use other operating systems/graphics cards, but I do not own any of them, so I will not pretend like I know how to set them up. I know Windows and NVidia. If you have questions about anything else; google it. I can not help you, and will not learn other platforms to do so. (Boy becomes a man, bro. Figure it the f*** out.)<h2 id="comfyui:-l5jnkx5j3">ComfyUI: </h2><ul><li>after downloading and installing<a rel="ugc" href="https://github.com/git-for-windows/git/releases/download/v2.47.1.windows.1/Git-2.47.1-64-bit.exe"> Git For Windows</a> (default settings) | </li><li>as well as <a rel="ugc" href="https://www.python.org/ftp/python/3.10.11/python-3.10.11-amd64.exe">Python 3.10.11</a> (default settings EXCEPT CHECK THE "ADD TO PATH" BOX ON THE FIRST SCREEN the rest default) | </li><li>make a folder called 'AI' in C:\ | </li><li>then open command prompt (win key+R type 'cmd' hit enter) | </li><li>then type 'cd /d C:\AI' and hit enter | </li><li>then 'git clone <a target="_blank" rel="ugc" href="https://github.com/comfyanonymous/comfyui">https://github.com/comfyanonymous/comfyui</a>' enter | </li><li>then 'cd comfyui' enter | </li><li>then 'python -m venv venv' enter | </li><li>then 'venv\scripts\activate.bat' enter | </li><li>then 'pip install torch torchvision --index-url <a target="_blank" rel="ugc" href="https://download.pytorch.org/whl/cu121">https://download.pytorch.org/whl/cu121</a>' enter | </li><li>then 'pip install -r requirements.txt' enter | </li><li>then download this model, and move the file to 'C:\AI\comfyui\models\LoRAs" |</li><li>then take your flux-dev.1 safetensors file, and put it in 'C:\AI\comfyui\models\checkpoints" |</li><li>then take your ae.safetensors file, and put it in 'C:\AI\comfyui\models\VAE" |</li><li>then take your clip-l and t5xxl files, and put the clip in the "clip" folder, and the t5xxl file in the 'text_encoder' folder located at 'C:\AI\comfyui\models\" |</li><li>then, back in the command prompt, type '<a target="_blank" rel="ugc" href="http://main.py">main.py</a>' enter &lt;---- run this from the "c:\ai\comfyui" folder to start comfy each time thereafter to start comfy when you want to use comfy |</li><li>then add the following nodes by double clicking on the background of comfyui in your browser, and then typing the following items in, one at a time, and searching them in the window that pops up, making sure they appear in to your comfy environment (I recommend adding all the nodes below, then going back and following the instructions for each one. <ul><li>"Checkpoint Loader", which you'll click the dropdown and select your Flux-dev model, then drag the connection on the right hand side/node output to the input of the 'LoadLora' node.</li><li>"Dual Clip Loader", and select 'clip_l.safetensors' on the first, and 't5xxl+fp16.safetensors' on the second. Then connect the "Clip" output to the Input of the "Load LoRA" node.</li><li>"Lora Loader", and after connecting the "model" output from "Checkpoint Loader" into the input for the Lora Loader, connect the CLIP from dual clip loader output node into the lora loader input. Take the "model" and "clip" outputs on the right side of the node and connect the "model" output to the "model" input on the "KSampler" node, and connect the "clip" output to the "Clip" input of the "CLIP Text Encoder" nodes. You can use the same Clip output on the Lora Loader node to connect to BOTH Clip text encoder inputs.</li><li>"VAE Loader", and select 'ae,safetensors', and connect the output to the input of the 'VAE Decode' node.</li><li>"KSampler", and set 'control after generate' to random for different pictures each time, and the steps to between 22-32, the CFG to 1.0 (although it'll work up to about 3.5 or 4.5, but past that its pretty shitty. i keep it at around 2.4), the "sampler name" to "euler" (not many others even work with flux models), and select the 'scheduler' to "sgm_uniform"... same deal, not many others work with flux, but you can try if you're bored.</li><li>"Empty Latent Image", and select '1024' for both width and height, and connect the "latent" output to the "latent_image" input of the "KSampler" node.</li><li>"Clip Text Encoder", and connect the 'clip' output from the 'dual clip loader' node into the 'clip' i</li><li>"Clip Text Encoder"</li><li>"VAE Decoder", and connect the output of the "VAE Loader" node to the input. Connect the output from the 'EmptyLatentImage' node to the 'Samples' input. Then connect the 'image' output node to the input of the 'save image' node.</li></ul></li></ul>

Files with SHA256: 0103e8870406d7f8ec4575f88455249323f2ac2b7acf2fc3c57f73e43af3365d