<p>This one's for foot fucking fans. SD really doesn't want to make normal-looking feet or put dongs between them, but this model is ready to get the job done. Can need decent weight but can still coexist okay with character LoRAs. Trained on all real photos but can make great images in any style, on any model that knows what a penis is.</p><h3>TL;DR</h3><p>Start with ChilloutMix, <code>a woman giving a footjob</code>, and <code>bad-picture-chill-75v</code>in your negative and add stuff from there. Weight 0.8 and CFG 7 is a good starting place.</p><h3>Main concepts/triggers</h3><p>Above all, the whole dataset is POV. You can use<code>pov footjob</code>, though synonyms like <code>a lady what with no clothes on is rubbin her foots on some dude's penis</code> apparently work just as well.</p><p>Understands <code>barefoot</code>, <code>stockings</code>, <code>fishnet stockings</code>, <code>latex stockings</code>. Easier to get normal-looking feet when the toes are hidden inside stockings! Also <code>pantyhose</code>, <code>sheer</code>, <code>nylons</code>.</p><p>Knows standard <code>footjob</code> and <code>reverse footjob</code> positions. Also <code>vertical footjob</code> for that position where it's under one foot and over the other, but that one basically doesn't hit.</p><h3>Tag format &amp; suggested core prompt format</h3><p>The dataset was tagged manually with this format:</p><p><code>a woman wearing</code> outfit-goes-here <code>is</code> <code>sitting</code>/<code>laying</code> on a <code>bed</code>/<code>couch</code>/<code>floor</code> <code>giving a footjob</code> <code>with her toes</code>/<code>with her soles</code>, <code>pov footjob</code>, <code>man laying</code>/<code>kneeling</code>/<code>standing</code> <code>pov</code>,</p><p>...and then comma-separated automatic tags from wd-v1-4-swinv2-tagger-v2. A similar format works well for prompts. (It doesn't actually seem to care about soles vs. toes)</p><h3>Negative prompt</h3><p>Pretty much required to get normal-looking feet. Great results with<code>bad-picture-chill-75v</code>. Traditional negative prompts can also work but not as well, it seems.</p><h3>Known issues</h3><ul><li><p>Struggles to generate people of color, will fix.</p></li><li><p>Likes to add watermarks, will fix.</p></li><li><p>Likes to put a stitched seam at the end of latex stockings, probably some wetlook tagged as latex, will fix.</p></li></ul><p></p>

Real Footjobs POV

<p>Trained on <a rel="ugc" href="https://civitai.com/models/6424?modelVersionId=11732">ChilloutMix</a> for 400 impressions per image (25 repeats X 4 epochs X 4 gradient accumulate steps), batch 6, Lion optimizer, 0.00005 unet LR, 0.00002 tenc LR. 207 manually-tagged images in 6 categories. This was an intermediate save from the ~26th footjob model training test, and the first release.</p><p>This one still has watermarks in the corners of a few input images, which are easy to prompt out but should all be context-aware-filled by next release.</p>

Files with SHA256: 37b86c43f622a77371df62e991877ff25d683a866fefc3e79d74bede464aeef0