<h2>This model makes deepthroats <em>deeper</em>.</h2><p>Trained on 68 images of (both real and anime) deepthroats that go all the way to the base.</p><p>Works with both anime and realistic models (tested with AbyssOrangeMix3 and UberRealisticPornMerge). Works with large weights up to 1.0, generally getting deeper deepthroats the higher the weight (see the example image with the weights).</p><h3>Strengths:</h3><p><u>Keeps artstyle intact</u></p><p>While some LoRAs tend to change the visual style when applied with large weights, my LoRA keeps the style of the checkpoint model intact, even at a weight of 1. I'm not sure how, but mixing real life, anime and more cartoon-ish artstyles in my training images may have done the trick.</p><p><u>Rarely adds text or motion lines</u></p><p>As long as appropriate negative prompts are used, text, motion lines, floating symbols etc. are kept to a minimum. I made sure of this by selecting mostly training images where none of these elements are present.</p><p></p><h3>Limitations:</h3><p><u>No reverse fellatio</u></p><p>Can't do any reverse fellatio positions on its own. Seems to be restricted to vanilla kneeling and lying blowjob positions, despite the training images containing more positions.</p><p><u>Biased towards spitroasts</u></p><p>Likes to do spitroasts, even if related tags are inserted into the negative prompts. The tested models however also did this on their own without my LoRA active, so I don't think my model is making it worse, just not better either.</p><p></p><h3>Please, feel free to posts some generations here!</h3><p>The cover image (and images in the model weight comparison) used hires fix with the SwinIR_4x upscaler, the rest is untouched. I used the orangemix VAE for all images generated with AOM3.</p>

Files with SHA256: f3bdb15b3eee9bf31313739b9dbd62f36870e4aca8f4ab831630c45eb1f7cf99