<hr />V2: see "about this version" description.<hr /><h3 id="recommended-weight-is-less-now-start-with-0.6-cumgtty3d">Recommended weight is less now, start with 0.6</h3>I misspelled the word "glans". I used "gland". Sorry.Trigger Keyword:<pre><code>Bussy</code></pre>Supporting prompts:<pre><code>from behind; spreading ass; shaved ass; hairy ass; penis from behind; visible penis; big/small penis; gland (yes is misspelled, sorry); large penis; asshole; tight anus gap; open anus gap; testicles; scrotum; feet; soles; hands; looking back to the viewer</code></pre>Negative supporting prompts:<pre><code>colored regions, (red, magenta, green, yellow:1.2), Shaved/hairy ass, visible penis, big/large/small penis, (hands:1.3), tight/open gap, vagina, woman</code></pre>Because of the "Color-associated region" training technique. Using "colored regions" and the color itself on the negatives might help OR NOT the generation. I'm not sure yet. Needs more testing.Problems with the current Lora:<ul><li>Hands are great, but I think feet is probably worse in v2.</li><li>In a complex scene is really common for the penis to be morphed with another concept. Like it becomes a stick, a weapon, or a log. It's hard because it's how SDXL behaves, bleeding concepts. Just do a simpler prompt or do in-painting.</li><li>Might do ass close-ups even when you don't want. Just describe the scene more. Reference the feet, etc.</li><li>I think v2 bias do giant dicks is pretty much fixed. But it can happen sometimes and scrotum and perineum can be weird connected. But IMO this version is way better.</li><li>Nipples on the ass are really rare now.</li><li>Maybe too much hair. But I like ti, sorry. Negatives on that and "shaved" on the positive works.</li></ul>Some more settings: Trained 1024 res. 207 images. Epoch 38 out of 44. It had block analysis, reduced TE to 0.5 and is added a 0.25 merge of Bussy v1.<hr />V1 old description:This is a SDXL version of <a target="_blank" rel="ugc" href="https://civarchive.com/user/phantom_phreak">phantom_phreak</a> SD 1.5 <a target="_blank" rel="ugc" href="https://civarchive.com/models/190261/bussy-boy-pussy-realistic-male-ass-and-testicles-pose">Bussy</a>. Check that out!<a target="_blank" rel="ugc" href="https://civarchive.com/user/phantom_phreak">phantom_phreak </a>got me his dataset to train it for him for SDXL. It was trained months ago and I was waiting for him to publish it, but he didn't responded. So for now I'm publishing it myself, but if he wants to publish it himself I'll take it down.And yes it CAN make woman/trans, even though it had absolute no images on the dataset, it works by just mentioning it on the prompt.I did some modifications. Reviewed some captions and added more images (I might have drove the model a little bit more to my taste (hairier men), but it still can be controlled well. <h3 id="v1-recommended-weight-is-0.93-good-from-0.8-up-to-1.0-ybou5akyj">v1 Recommended weight is 0.93‎ / Good from 0.8 up to 1.0‎</h3>Start with 0.93 but if you need more flexibility you can lower it down to 0.65‎ ‎Trigger Keyword:<pre><code>Bussy</code></pre>Supporting prompts:<pre><code>from behind; spreading ass; shaved ass; hairy ass; upside-down penis from behind; visible penis; big penis; large penis; asshole; gaping anus; low-hanging testicles; scrotum; feet; soles; hands; looking back to the viewer</code></pre>Negative supporting prompts:<pre><code>Shaved ass; hairy ass; (upside-down penis from behind, visible penis, big penis, large penis:1.3), (hands:1.3)</code></pre> ‎It's hard to not get giant dicks, so if you just want testicles showing, do not mention "penis" on the positive prompt and put "(upside-down penis from behind, visible penis, big penis, large penis:1.3)" on the negative. It normally works. I changed "shaven" from the SD15 model to "shaved" here. Not sure what is the best, but this SDXL uses "shaved".‎Some good LoRa combinations I've tried:‎<a target="_blank" rel="ugc" href="https://civarchive.com/models/218698/realistic-male-ass">Realistic Male Ass</a>: at 0.3 sometimes gives better skin textures and hairy assholes, give more bubbly butts.<a target="_blank" rel="ugc" href="https://civarchive.com/models/279972?modelVersionId=328411">Experimental guided training of Color Associated Anatomy of the pussy</a>: I've experimented with this LoRa because it has knowledge of the anatomy of private parts of women AND men and their exact position. BUT you should use it at just 0.1 (don't go above 0.15 or you will lose quality as it was only trained for close-ups) and use on the NEGATIVE prompt:<pre><code>(woman, vagina, pussy, labia majora, labia minora, clitoris hood:1.1)</code></pre>‎edit: this still needs more testing, probably is not worth using it for male. Might erase this "suggestion".‎Other parameters and settings:‎The base checkpoint is the “sdXL_v10VAEFix” 6.7GB. So, it should be very flexible with any checkpoint.As of right now, I recommend <a target="_blank" rel="ugc" href="https://civarchive.com/models/133005/juggernaut-xl?modelVersionId=288982">juggernautXL_v8Rundiffusion</a> and <a target="_blank" rel="ugc" href="https://civarchive.com/models/245423?modelVersionId=288402">juggerxlInpaint_juggerInpaintV8</a> for inpaiting.‎And turbo <a target="_blank" rel="ugc" href="https://civarchive.com/models/112902?modelVersionId=333449">dreamshaperXL_v2TurboDpmppSDE</a> works quite well!! (remember you NEED to use DPM++ SDE Karras, CFG 2.5, 5-8 steps. I don't think the quality is the same, but it is good enough for a quick generation.My recommendations for turbo dreamshaper is to generate at 2.15 CFG with 8 steps and then highres fix for 0.45 denoise, 1.25x res, 6 steps, same checkpoint and same scheduler. Works great!‎For standard generationCFG: 5.5DPM++ 3M Exponential (50 steps or more)DPM++ 2M Karras (25 Steps or more) - this one looks like gives more of the "nipples on ass" problem, using FreeU v2 for SDXL looks good overall and fixes this for meDPM++ SDE KarrasDPM++ 2S a Karras‎‎I did not try Lora block weights yet. It might improve results.<a target="_blank" rel="ugc" href="https://github.com/cheald/sd-webui-loractl">Loractl </a>works great if you want to have a more complex prompt, loose the generation with 0.3 or 0.65 first and then increae it to 0.93 by the 20% step. Like this:<pre><code>&lt;LoraName:0.3@0.2,0.93@0.2&gt;</code></pre>‎Want to have some “fun”? Install wildcards dynamic prompts extension: <a target="_blank" rel="ugc" href="https://github.com/adieyal/sd-dynamic-prompts">https://github.com/adieyal/sd-dynamic-prompts</a> . And wildcard manager. Here is a prompt I made for testing. Paste on prompt:<pre><code>a __YetAnotherWildcardCollection-main/States/Nationality__ man {18|28|45|80} years old on a __YetAnotherWildcardCollection-main/Background/*__ laying spreading his {shaved|hairy|smooth|pimpled|bubble|squared} ass asshole from behind bussy, (gaping anus:1.2), {low-hanging testicles scrotum|upside-down penis from behind, big penis, large penis|upside-down (penis:-0.5) from behind, (small micro:1.5) (penis:-0.5)}, { |feet soles}, natural lighting, 4k, high quality, Fujifilm XT3 &lt;lora:bussy_SDXL:0.83&gt;</code></pre> ‎Problems with the current Lora:<ul><li>Hands and feet sometimes morph into other objects or are weird</li><li>Head position sometimes look bad</li><li>Bias to GIANT dicks from behind</li><li>Depending on the checkpoint, scheduler and number of steps you might get nipples on the ass. I use tags on the negativa to try to not get this like: "pimples, bruises, nipples"</li></ul>Some more settings: Trained 1024 res. 190 images. Epoch 22 as a continuation of a previous 15 epoch. now prodigy 1.0. 2 steps folder. constant BATCH 2, rank 64/1,Scale weight norms 1, snr gamma 5, Noise offset 0.0357, no regularization imageHopefully you can leave some results and some comments. Any idea is appreciated. Thank you.

Bussy - Boy Pussy / Male Ass and Testicles Pose from Behind – SDXL

This version is a BIG leap in quality. It improves a lot of the ass details plus anatomy comprehension. Improves:1-prompting in and out the visibility of the penis. 2-penis size. 3-the appearance of the testicles, glans (it is misspelled as "gland" in the dataset sorry, try using both), the anus gap. 4-gap size (tight/open), but still needs more training. 5-Ass being shaved or hairy. But it defaults to be a lot more hairier than v1.This still uses the trigger word "bussy" but works without it if you reference a person ass "from behind". Start with weight 0.6 up to 0.9.Training had 10% of Color-associated region images method to teach: asshole, anus gap, testicles, penis and gland. Trained version v4, epoch 38 out of 44. The final version is a merge with some block analysis (lbw=0.5,0.15,0.15,1.25,0.8,1,1,0.5,1,0,0.4,0) of v2+v1 &gt; 0.85+0.25.Next version I'm moving out of the trigger word "bussy". It was just a homage to PhantonPhreak but the LoRA not only does not need it but i think it bleeds the word "pussy" into the generation. Using "woman, vagina" in the negative changes details a lot, so I think even when it's not clear, the word does bleed.I hope I can make it more flexible as well with more control on the gap, maybe insertions, more interactions with other people (rimming, touching etc) and poses. As of right now, this is still a ass model with a fixed hardcoded pose.

Files with SHA256: 14643cde5e85f12d6f2407d959562e646a55ff96c8358721b0d3a96776715849