<h3 id="this-was-made-mainly-for-slime-girl-penetration.-i-have-not-tested-other-cases-where-transparency-would-make-sense.">This was made mainly for slime girl penetration. I have not tested other cases where transparency would make sense.</h3><h3 id="+qwen-image-2.1">+Qwen-Image-2.1</h3>Trained for T2I - Maybe editing works but I haven't tested it.I rewrote the dataset captions in natural language this time since Qwen-Image-2.1 wasn't quite as easy to train. I could only really train on resolutions of 1024-1536 due to VRAM limitations so it might not work or it might introduce unwanted anatomy issues over ~2.5 MP. Otherwise the model is pretty particular about prompting. If anything is misinterpreted or conflicting or you try to do something that nether the base model nor this lora knows about, you might get some pretty bad anatomy distortions (body horror). Some people are also confused about the generation settings for T2I but I usually do 40 steps with euler/simple or res_multistep/simple at CFG 3.0 and weight 1.0. I haven't really tested with turbo but if something is getting messed up try lowering the weight. Also this one can kinda work as a standalone lora but I would still recommend a helper lora or checkpoint for better and more consistent NSFW knowledge especially at high resolutions.<h3 id="+krea2:">+krea2:</h3>The dataset is the same as v5 for Illustrious which honestly worked better than expected (since the captions are still all booru tags). Krea2 doesn't have a good understanding NSFW stuff so you might need other LoRAs to help with certain concepts and positions. You don't need any specific bypass nodes or LoRAs to "unlock" Krea2, the LoRA(s) just need(s) to be good enough (and this one does seem to be good enough for that). I also haven't had enough time to do a lot of testing with Krea2 yet so there could be other issues I'm not aware of. Euler + beta works well for 8 step turbo generations.<h3 id="v5-update:">v5 update:</h3>I went through and fixed most of the transparency &amp; anatomy issues still present in the dataset. So now transparency should be better, make less patchy slime, and work for more things (like fisting, fingering, multiple penetrations). Note that there are still issues with high/from above angles, size differences for oral, and the orientation for some unique penetration types. Using a big stack of loras and/or a lot of high tag weights will also make proper transparency significantly harder to get.Getting proper transparency to work with slime girl penetration is a pain. Most approaches either use x-ray (which looks weird for slime girls), do some complicated prompting workarounds, or just deal with low annoyingly low success rates.The dataset is a mix of different resolutions, aspect ratios, slime colors, positions, and penetration types. It should be pretty responsive to color and transparency in most cases. You may need other loras to get concepts such as fisting, fingering, multiple penetration, etc. properly consistent though. <code>translucent penetration</code> should be the only thing needed to trigger it, but you might get better results with additional tags/phrases such as <code>slime girl</code>, <code>(object) in slime</code> <code>deep penetration</code>, <code>perspective</code>, <code>goo girl</code>, <code>monster girl</code>, <code>(color) skin</code> (such as <code>blue skin</code>, <code>green skin</code>, <code>multicolored skin</code>, etc), <code>(object) visible through body</code>, etc. Some penetration types might work better with specific angles (especially the types that are still WIP). If you are getting mismatched internal penis sizes you can try adding <code>(disconnected penis:1.2)</code> to your negative prompt.I recommend pairing this with a dedicated slime girl LoRA (such as <a target="_blank" rel="ugc" href="https://civarchive.com/models/1047424/dasiwa-illustrious-slime-girl">DaSiWa-Illustrious-Slime-Girl</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/2430224/realistic-slime-girls">Realistic Slime Girls</a>, or <a target="_blank" rel="ugc" href="https://civarchive.com/models/344495/slime-girls-concept">Slime Girls - Concept</a>) and a checkpoint that is able to handle NSFW concepts well such as <a target="_blank" rel="ugc" href="https://civarchive.com/models/1034922/dasiwa-illustrious-or-anime">DaSiWa Illustrious | Anime</a>. While a strength of 1.0 should be safe, it's possible the image could end up overcooked with other LoRAs. Try setting it to ~0.7-0.8 if you are getting distorted faces and/or patchy slime. That should be enough to guarantee the concept in most cases, but a good LoRA combination with some good prompting can get it to work with as low as ~0.25. My typical generation args are either:<pre><code>steps: 37
cfg: 4.0
sampler: euler ancestral
scheduler: simple</code></pre>or:<pre><code>steps: 40
cfg: 4.0
sampler: uni_pc
scheduler: ddim_uniform</code></pre>unless I have <a target="_blank" rel="ugc" href="https://github.com/Extraltodeus/Skimmed_CFG">Config Skimming</a> enabled then I set my cfg to 16 or 24 with the skimming cfg at 4 or 3 respectively.If you are having issues with four or six fingers and you are still using v4 or earlier, try updating to v5 and reinforcing your negative prompt.

Translucent Penetration

Still the same-ish v5 dataset, but the captions were all rewritten in natural language. The dataset will probably be reworked for future versions to fix some lingering issues<ul><li>Trained for text-to-image for Qwen-Image-2.1.</li><li>Can work as a standalone lora, but using a general NSFW lora would still help with NSFW positions, actions, and visual consistency.</li><li>Highly dependent on prompting (Qwen-Image-2.1 can be particular this).</li><li>Works best within 1.0 MP - 2.5 MP. 3.0+ MP can work sometimes but I wasn't able to train that high.</li></ul>Use with CFG ~3.0 and a weight of 1.0 unless you are using a turbo lora

Files with SHA256: a7fca96a27d41b75770d43a3a88ee9a07c1ab785b07f8f0543e83bd8c289f642