Hassaku_SD1-5_v1-3

Hassaku (SD1.5)_v1.3

Description:

<p><strong>Small update on those things:</strong> </p><ul><li><p>Slightly bigger breasts in compare on prompts large and huge breasts on some images</p></li><li><p>Slightly better formed hands midrange (from bad hands to slightly better bad hands... on some images it helps)</p></li><li><p>Better response on prompt fellatio with a higher probability, that the tip is in the mouth</p></li><li><p>Details on clothes or bodys (like wings) are less likely to be detached from bodys and is more equal right and left</p></li><li><p>Less extra limbs, ears, animal ears, wings</p></li><li><p>Seems that details on clothes are less random formed out</p></li><li><p>Slightly better constructed backgrounds</p></li></ul><p> <strong>Hard to determine:</strong> </p><ul><li><p>Waists seems more norrow, not sure about that</p></li></ul><p> <strong>What is worse:</strong> </p><ul><li><p>On some prompts, clothes apear more lose with more "strings" or holes all over the place<br /><br />Be aware that those points are tested over many images on clip2, it can be, that on some images, those points don't do anything or are even worse than before. It is hard to update things with training, if most training material don't look like the images my model itself produce. Need to merge back and forth, so the model itself looks and behaves like before training. It have less changes with training, because of it. Model trained on 150 fellatio images and later merged back with hassaku and a bit of koji. Archived a bit more stability with add difference (on lower resolution).</p></li></ul>

Civitai Page: https://civitai.com/models/2583?modelVersionId=106922