This is the first Pony version of my inverted Oral LoRA. I trained this one on over 200 images from various styles and made an effort to revise deficiencies in the source, but these models struggle with drawing upside down ears I guess, especially when the training data is not pure perfect ears and fingers. Of note, this model is less likely to insist on side view than V0.9 and likewise has considerable training data that includes gangbang scenarios. Check the example images for prompt examples.<h3 id="selecting-a-pony-checkpoint">Selecting a Pony Checkpoint</h3>I thought it was just me but this checkpoint is a bit touchy perhaps. After some testing, I'm going to announce that for most use-cases, the only checkpoints with greater than 50% reliability per random seed once you home in on a good balance of prompting and tuning the LoRA strength, are Pony Diffusion, Crody's Pony 3DCG XL 2.0, and Blendermix -- these will have a 50% shot at successfully rendering something that at least looks like inverted oral. Most other models omit the male, despite being prompted for. If the male does not show up, try raising the CFG.General Prompt Guidance CFG greater than 3.5 recommended masterpiece, best quality, amazing quality, 4k, very aesthetic, high resolution, ultra-detailed, absurdres, newest, scenery,1girl, ((her lora name, any descriptive details)) solo focus, on back, (head upside down, head over the edge,) (on bed,) erect nipples, inverted fellatio, throat bulge, medium angle, {saliva, cum on face, cum on floor,/ } ] ,{2boys, gangbang, missionary sex,/1boy,} erect penis, {hand on throat,/ }night, lamplight, bedroom, lamp, window, plant, trees, fog, moon,Example promptscore_9_up, score_9, score_8_up, score_8, score_7_up, score_7, (side view:1.3), raw, high quality, 4k resolution, absurdres, 1girl, 1boy, solo focus, ((blonde, closed eyes, medium hair, nude, )) solo focus, (side view:1.4), head and shoulders shot, on bed, on back, (head upside down, head over the edge,) (medium breasts,) (inverted fellatio:1.2,) 1boy, hand on neck, head out of frame,hotel room, window, curtains, trees,side view, close up,<h3 id="zeroing-in-on-consistent-results">Zeroing in on Consistent Results</h3>This LoRA's training includes some "deepthroat" and "gangbang" keywords. The latter should work reliably. This LoRA does not work well combining Character LoRAs with Morimee and Incase styles, even at weaker strengths, it induces artifacts.Sometimes embeddings can break the LoRA, depending on what checkpoint you are using and how far outside the typical parameters (on bed, on sofa) you go. If you are combining style and character LoRAs and then throwing embeddings on top, you're even more likely to be met with disappointing results. Maybe a 2.0 will be able to overcome that.Unlike 0.9, this LoRA is not trained on the following because it didn't really yield results so maybe these keywords are best left for general use.side view, close up, medium angleBut using these, and adding weight (1.3?) to these, can overcome misinterpretations and conflicts!For example, if close up and thighhighs are prompted, the model will become confused and most likely shoot for a medium angle with the legs up in the air -- not a close up.So instead for example, add weights likeside view:1.3, close up 1.3,That will clear things up for the LoRA and is easier than pondering the details of every prompt. The LoRA is robust enough to provide results, even if the prompt is conflicted. The training data was generated from stills and a scattering of video captures, using my LorAs and previous epochs AND using the <a target="_blank" rel="ugc" href="https://civarchive.com/models/461245/over-the-edge-blowjob">Over the Edge Blowjob</a> LoRA from @freckledvixon for a better variety or to overcome bias issues with my model and my 0.9 model’s overtraining.This model still has some 3DCG bias, so for anime or realism, increase the strength and/or use fewer LoRAs that might conflict (such as character LoRAs and styles). On-site generation might disappoint when using Realism (it's locked in for 3DCG/Anime), if you are careful with gold buzz expenditures then I recommend NOT using this model with realism, the results won't get you far. Don't even try unless you can do so freely.

Files with SHA256: 59629248f310fd8ffe5158b6a7ea44b4d7ae01897f43e638a0da59fd18e612aa