<h2 id="concept">Concept</h2><p>Edit an existing image of a person to be making a JOI gesture with one of their hands. Never built an edit LORA before and I have no idea what I'm doing. So take the lora with a grain of salt and hopefully it works.</p><h2 id="usage">Usage</h2><p>I'd start with 1.0 strength. If hands are visible in the frame it's usually does better. It's almost entirely been trained on right handed images because finding left handed people is hard. Best images are ones where at least the upper body is visible. Images which are headshots only tend to be hit and miss.</p><p></p><h3 id="base-prompt:">Base prompt:</h3><pre><code>Adjust the woman's right/left hand to make a JOI gesture.</code></pre><h3 id="hand-placement:">Hand placement:</h3><pre><code>Position her right/left hand near her chest/crotch/shoulder/buttocks/face.</code></pre><p></p><h2 id="training">Training</h2><p>The model was trained on 25 image pairs grabbed from videos. Trained at 1mp for 8 epochs on a single 3090 ti, using Fizgig. Captions were applied directly in fizgig, with some manual updating for hand positions.</p><p></p>

Files with SHA256: 4cdd3d1f61463cf345e4cdfc29596802f6afa7460590a8d0b356d504033c6fca