<span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:16px">Want to support my work or help fund the training of this dataset on other models? Join the Patreon in my profile, and if you do - thank you!<h3 id="krea-2-v1.3d:">Krea 2 V1.3D:</h3>Alright, alright. I figured out a way to greatly restore texture back. It wasn't more training or different settings, and that's all I'm going to say about it. I deserve to have some trade secrets, no? I made it version 1.3D because the D stands for detail. It was actually try F, as in I was totally...frustrated...with how many lengthy tries it took.Anyways, I had versions that took it further but it tended to have more broken anatomy. So, the same rules still apply if you want something to look like a photo:Be sure to mention that it's a photo/photograph in the prompt. Do NOT use "photorealistic" or any other term that doesn't actually mean a photo. Krea 2 was trained on a ton of artwork so using any of those types of terms will yeet it back to bad texture.Some rare prompts might still want to go over to the anime-side, but I still have the photo slider directly on SNOFS Krea. Simply set it to somewhere around .5-2 on the strength and you're probably golden. This was trained on SNOFS 1.2 so I'll be putting out an updated version at some point: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2823820/photodetail-slider-for-snofs-krea">https://civarchive.com/models/2823820/photodetail-slider-for-snofs-krea</a>The same goes for some specific terms. When in doubt, use the words listed below instead of their synonyms.My two-stage sampler is still handy for multiple reasons: <a target="_blank" rel="ugc" href="https://github.com/Auryg/Krea-2-Two-Stage-Sampler">https://github.com/Auryg/Krea-2-Two-Stage-Sampler</a>The idea for the two-stage sampler that you can generate without the turbo lora for the first bit, which helps with variation and prompt adherence, and then bump to a second stage with the turbo lora to keep things fast. For the three-stage variation, you can then bump back to doing it without the turbo lora for the last bit if you want to use a negative prompt. You can also generate at a lower resolution to start and then bump it up.Also, please, for the love of god try SNOFS by itself before you go adding a bunch of other general NSFW loras or models to it that screw up anatomy.<h3 id="ideogram:">Ideogram:</h3>Ideogram model has been updated, and is available here: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2781404/sex-nudes-other-fun-stuff-ideogram-snofs">https://civarchive.com/models/2781404/sex-nudes-other-fun-stuff-ideogram-snofs</a><h2 id="general-information:">General Information:</h2>SNOFS was trained on natural language (or JSON, for Ideogram), not tags. It will work best if you use full sentences to describe what you want.Not using ComfyUI/your inference software doesn't support lokr? I've put up a merged version here. You can also use the merged base model to train off of: <a target="_blank" rel="ugc" href="https://civarchive.com/models/2416142/snofs-sex-nudes-and-other-fun-stuff-merged">https://civarchive.com/models/2416142/snofs-sex-nudes-and-other-fun-stuff-flux-2-klein-9b-base-and-distilled</a>Here's a list of some of the terms that work well:<ul><li>anus</li><li>blowjob</li><li>boudoir</li><li>condoms</li><li>deepthroat</li><li>braless</li><li>cowgirl position</li><li>cum</li><li>cunnilingus (be specific and maybe put kissing in the negative prompt)</li><li>deepthroat</li><li>dildo</li><li>doggystyle position</li><li>fingering (anal and vaginal)</li><li>hand in panties</li><li>handjob</li><li>hitachi magic wand</li><li>implied blowjob</li><li>ipcam / nightvision ipcam</li><li>masturbating (might want to put penis in negative prompt, or specify what she's rubbing for women)</li><li>massage</li><li>missionary position</li><li>naked, nude, etc.</li><li>penis</li><li>pregnant (and can specify trimester)</li><li>prone position</li><li>reverse cowgirl position</li><li>sex</li><li>sheer</li><li>snapchat (and caption/text/etc)</li><li>selfie (and mirror selfie)</li><li>spooning position</li><li>strap-on dildo</li><li>tentacles</li><li>licking testicles</li><li>undressing</li><li>vagina</li><li>wet clothes</li></ul>Depending on the version, the following might work:<ul><li>anal sex</li><li>anilingus</li></ul>But also keep in mind that it was trained on stuff like "her panties are pulled down to her thighs," not "panty pull."These models are under the following license:<a target="_blank" rel="ugc" href="https://huggingface.co/Ashen3/SNOFS">https://huggingface.co/Ashen3/SNOFS</a><h3 id="flux-2-klein-9b-v1.4:">Flux 2 Klein 9b V1.4:</h3>Additional training. Some of the training was done using <a target="_blank" rel="ugc" href="https://github.com/BuffaloBuffaloBuffaloBuffalo/ai-toolkit-perceptual">https://github.com/BuffaloBuffaloBuffaloBuffalo/ai-toolkit-perceptual</a> , training against depth. Considering how much of SNOFS is two people intermingled with close skin colors, it seemed like a novel idea. It did seem to rapidly help with that sort of thing. On the downside, it seemed to create a bit of a texture issue on very close up images. I did some more training after to try to bring that back and was somewhat successful, but I think I'd need to increase the weight decay to really make that happen. Since everything else was in a good state I decided to release as-is. If you do have that texture issue, try adding "goosebumps" as a negative prompt.

SNOFS (Sex, Nudes, Other Fun Stuff)

qwen_snofs.safetensors

v1.0

Krea 2 v1.3D

Krea 2 v1.3

Krea 2 v1.2

Krea 2 v1.1

Krea 2 v1.0

Ideogram v1.0

Klein 9b v1.4

Klein 9b v1.3

Klein 9b v1.2

Klein 9b v1.1

Klein 9b v1.0

Qwen v1.3

v1.2

v1.1

Files with SHA256: 209c366827cf432f691eea95abdd339aa631ad7037aacb9203237541197b4deb