Redefines most "testicles" to be softball sized at the very least. My goal was to allow for bigger testicle sizes and more contrast between testicle and penis sizes since most futanari-centric models are already primed for huge and hyper-sized penises."huge testicles" is intended to make the combined size of the testicles as large as the character."gigantic testicles" is intended to make them larger than the character, possibly going off the frame.Of course, YMMV with the model, lora weight and emphasis used. Trained on NAI animefull, lower the lora weight if it's frying on your merge.Has penis size tagging so that umprompted penis size will be strictly smaller than the testicles, but you can try for larger sizes. "small penis", "huge penis" and "hyper penis" were used.Try the following:<ul><li>"sagging testicles" - self explanatory</li><li>"testicles on ground", "testicles on floor" - Should help shape testicles resting on a surface (typically the ground)</li><li>"squishyballs" - tag used on pictures where the testicles are visibly deformed from being pressed against something (a limb, a surface, the ground, etc)</li><li>"balls behind" - tag used on pictures where the testicles were located behind the thighs.</li><li>"self sitting", "sitting on balls" - Ow! My Balls! Primarily used on pictures with "gigantic testicles".</li></ul>"hyperballs" used to be the activation word common to the entire dataset but I'm having better results without it.Version 4<ul><li>Dataset expansion, courtesy of the<a target="_blank" rel="ugc" href="https://civarchive.com/models/310084"> Pony Lora.</a>, primarily including more gigantic testicles.</li><li>"backsack": The stretch of foreskin that joins the testicles with the groin, visible from behind. Slight overlap with "sagging testicles". Can be used with "perineum","pussy" or "vulva".</li><li>Further increase in resolution, up to 896.</li></ul>Version 3<ul><li>New GPU means vastly increased training resolution and dim, translating into much higher quality.</li><li>More support for "gigantic testicles"</li></ul>Version 2_3<ul><li>Trained for significantly more steps (~3x)</li><li>Significant bump in dropout and scale weight norms. Should burn up less on other checkpoints.</li><li>More exhaustive tagging, resulting in better discretization of testicle sizes.</li><li>Token increase in training set (~15 images) to patch up blind spots</li><li>Downsized with sv_ratio 20. 25% of the size, 99.89% of the data</li><li>"squishyballs" - Experimental tag which was used on pictures where the testicles are visibly deformed from being pressed against something (a limb, a surface, the ground, etc). Too few images for it to have a significant impact but I found it lessens the instances of melding legs.</li><li>"balls behind" - Experimental tag to specify that the testicles should be located behind the thighs. Too gacha to be reliable but it's there.</li></ul>Like its previous iteration, it works best on <a target="_blank" rel="ugc" href="https://civarchive.com/models/110541/futanari-factor">Futanari Alpha</a>. It also found moderate success on <a target="_blank" rel="ugc" href="https://civarchive.com/models/31546/based65">Based65</a>, and surprisingly <a target="_blank" rel="ugc" href="https://civarchive.com/models/8281/perfect-world">Perfect World</a> (despite being an ostensibly anime model). Works at 1.0, you might need to lower it on less compatible checkpoints and emphasize keywords instead.Issues<ul><li>Still not as widely applicable as I'd like across different models, but it's definitely a step up.</li><li>Scuffs faces real bad. It is not a problem for me since my routine involves upscaling from resolutions adjacent to 576x576, which fixes it for the most part, but its something to keep in mind. I recommend halving the strength you're using before upscaling.</li><li>Some models such as the AbyssOrange mixes have a tight stranglehold on "testicles", and it might not have an effect there. You might have to use "huge testicles" instead and play with emphasis.</li></ul>

Hyperballs

<ul><li>Used weighed training to diminish the effect of dataset noise and improve anatomy.</li><li>Added a few more "self sitting" examples to the dataset to address popular demand. It is still not very reliable at it, but should be easier now</li><li>Modified some "testicles on floor" and "testicles on ground" examples to better shape the testicles when they are resting on the ground, causing them to slightly flatten under their own weight. This can help avoid the "two orbs" look when prompting those tags at high weights. </li></ul>

Files with SHA256: 2b20ccc54d6c0965db371853b8bde3c2fe6b2281bc07d5133c3cb68d8468ca58