Ultimate Cunnilingus Concept Universal (Heterosexual) SDXL

This is V1 of the Ultimate Cunnilingus Concept Universal (Heterosexual) for SD XL. I created this in order to easily evoke cunnilingus on most SD XL checkpoints without having to resort to Pony XL.It utilizes the same, huge dataset (3K+ training images) as the SD 1.5 versions, making it quite possibly the largest cunnilingus LORA for SD XL.In addition, it follows the same concept and prompting as the SD 1.5 versions, so if you are familiar with those then it is essentially the same here. A bit more consistent than the SD 1.5 versions, so should require less generations to get good results (on the contrary, SD 1.5 versions still produces the best photorealism results imo).This is universal in style, so it will work on photorealism, anime and in-between checkpoints.Remember to use ADetailer or img2img inpainting to fix faces (see ADetailer section).Currently in development and will definitely be improved upon. Let me know of any issues or things you like added. Training only 4 epochs took over 70 hours on a 4090, so it may take some time for future versions!Requires LyCORIS support (not a problem if your GUI is up to date).Compatible with SD XL checkpoints only.Usage:No unique trigger word is necessary, but you should follow the example prompts for consistent and best results.Weight of 0.7-1 is recommended (0.8 being the sweet spot). If the weight is too high, faces can become distorted (using ADetailer can help - see below). Lower weights will allow the checkpoint more expressive freedom in generating cunnilingus, but higher weights will generate more consistent results.CFG scale should be 2-4. I find that 2-4 for XL gives the best quality yet still follows the prompt very precisely. There isn't really a need to go above 4 imo. There is a big risk of introducing artifacts at higher CFG scale.Recommended base resolutions are: 1024 x 1024 (square), 640 x 1024 (portrait) and 1024 x 640 (landscape). These are from a lot of experimentation! I find that portrait resolutions gives the best and most pleasing results. It is not recommended going higher unless using HiRes fix or img2img upscaling.Sampling method: DPM++ 2M for photorealism and Euler a for anime. Automatic scheduler for both.Use at least 30 sampling steps!Use clip skip 1 regardless of checkpoint.Make sure you use an appropriate VAE. I use this one.Example Prompt:Most Basic Prompt:(((1girl, 1boy, hetero, cunnilingus, boy licking her vagina, boy licking girl's vagina, boy licking her pussy, boy licking girl's pussy, background blur, best quality, masterpiece)))You should always add more specifics however as it can result in less weird and randomness, such as setting, breast size, hair color, clothing parts, etc.Optional tags that may provide more varied results:hand on another's head, spread legs, ass, anilingus, sitting on face, clothes lift, full body, from above, from below, from front, from side, open mouth, closed eyes, tongue out, pubic hair, pussy juice, pussy cumNo negatives should be necessary, as SD XL is quite good at default with none.Generation:It may take a few attempts at generation to get a good result. Semi-good results can be salvaged with img2img generation or inpainting. As with any concept LORA, playing with RNG is a part of the game (usually with the player losing awfully). If you do not get good results or the checkpoint is not generating cunnilingus, keep trying! If after about 10 generations (considering your prompt follows the example + using the correct sampling method) and nothing is turning out right, definitely use a different checkpoint. You can look at my example images for the checkpoints I use and recommend.ADetailer is HIGHLY ADVISED!Using ADetailer with detection model confidence threshold of 0.83 will almost always exclusively detect the female face and fix it, making it look much better. Males usually do not need to be fixed, and fixing them with ADetailer can cause them to look messed up (that's why a high detection rate is necessary).Make sure to additionally put this into the ADetailer prompt: 1girl, cunnilingus, detailed faces, best quality, masterpieceDo NOT put this LORA in the ADetailer prompt for best quality!Use inpaint denoising strength of 0.4-0.6.Alternatively, you can also fix the faces later in img2img inpainting (using the same ADetailer prompt above). ADetailer however is just a lot quicker with less steps.Can I use this for yuri/lesbian cunnilingus?You can try, but I have not attempted this nor do any of the trained images are of yuri/lesbian. If this does not work well, there are other existing yuri/lesbian cunnilingus LORAs that already do this and to a good extent. However, in future versions I may additionally include support for yuri/lesbian cunnilingus.Can I use this in combination with other LORAs like characters?Yes, but be careful with using too high of a weight on the character LORA or too many LORAs at the same time. I recommend one character LORA only with weight of 0.7-0.8 (on the character LORA).Is there any POV cunnilingus?Not at the moment. I'm not a big fan of POV style and there are already plenty of POV cunnilingus LORAs available. If you want POV added, let me know and I can try adding it on future versions. However, adding POV will make it more difficult to generate non-POV images, despite correctly tagging with triggers (as exemplified with the many universal cunnilingus LORAs available). Therefore, a separate POV only cunnilingus LORA may be necessary.Can I choose the specific position of cunnilingus to generate (e.g. standing, sitting, lying, etc.)? Not yet precisely, as this requires separate trigger tags for every position and is very difficult and time consuming to implement (perhaps in future versions?). You can however try adding position tags and see if it helps, but there is no guarantee SD will follow it. At the moment, it's best to be vague about position and let SD roll the dice so you can get the most unique results.The results look REALLY BAD or I'm not getting cunnilingus!First, make sure you are following the example prompt exactly. Second, make sure you are using the correct sampling method (DPM++ 2M for photorealism and Euler a for anime). Afterwards, if you're not getting cunnilingus, just keep trying to generate and you'll likely get something as it can take a few generation attempts (this is just due to the current nature of SD XL not really understanding the concept very well).If after about 10 generations and you're still not getting cunnilingus or just bad results, it's probably an incompatible checkpoint. Some checkpoints (most notably non-NSFW ones) do not work well with this LORA or at all, unfortunately. It's not possible to make sure it looks and works well on every possible checkpoint, so please try a different checkpoint. Again, please check my example images for recommended checkpoints to use.You can also try lowering the weight if it's simply an issue of image quality.Check back often for updates!This is a serious project for me and I will be updating, fixing and improving this LORA as often as possible, including adding new training data for even more delicious cunnilingus. So please bookmark and check back often for new updates!Feedback welcomed!I'd love to hear you feedback and ways I can improve this LORA. Sharing with others is caring, and is the only way I have the motivation to release and update LORAs for free!