<p>This is an experimental LoRA, not super polished. It often takes me several attempts to get a decent result! But I'm hoping you guys can have fun with this.</p><p></p><p>This LoRA puts you in a world where you possess a magic phone that shows you the people around you without their clothing.</p><p></p><p>To use the LoRA, place the <code>majikfone</code> keyword at the beginning of your prompt. Hope you enjoy!</p><h1 id="heading-1165">Versions</h1><h2 id="heading-1166">v1.0</h2><p>This version has a better grasp of how the hand holding the phone should look. However, for certain prompts it will produce "overbaked" output. You may be able to correct this by a combination of lowering the LoRA weight (the sweet spot is usually somewhere between 0.8 and 1.0) and decreasing the CFG scale (between 4 and 7). It also seems to work most consistently on 512x512 square images.</p><p>In general, if your output image has multiple phone screens, weird winding phone borders, or blatant nudity (with no phone at all), then lowering the LoRA weight may help. On the other hand, if you're seeing clothes inside the "phone" or the subject's nudity bleeds out beyond the phone edges, you may need to increase the LoRA weight. A lack of phone can be the result of weight being either too high or too low (and it might just be a bad seed).</p><h2 id="heading-1167">v0.1</h2><p>This was my first attempt at training this LoRA for SD1.5. I wasn't sure it'd be able to handle such a multifaceted concept. It tends to get very confused by the hand and fingers holding the phone, but it works correctly occasionally.</p>

Majikfone for SD1.5

<p>Increased the quantity and variety of training data in the hopes of making it more reliable.</p><p>I adjusted the block weights a bit after training to rescue it from overbaking effects, but you may still need to reduce the LoRA weight (0.8-1.0) or lower the CFG scale (4-7) for best results. Most reliable results seem to come at 512x512 base resolution.</p>

Files with SHA256: 79088663eb3eb8622fc21b7d603a57051c010b54a294ed62792a0f6d2c0bab30