<h1 id="raora-panthera-ryed2opsp">Raora Panthera</h1><h2 id="disclaimer-wwht035u7">Disclaimer</h2><ul><li>Do not use my LoRA to produce AI image and tagging her fanart hashtag in Twitter/X or if you have an enough contribution in that image then fine.</li></ul><h2 id="trigger-word-obnuom77q">trigger word</h2>debut costume<pre><code>raora1st, panther girl, animal ears, animal ear fluff, blush, pink hair, streaked hair, large breasts, long hair, yellow eyes, goggles, black tube top, black hat, white thighhighs, cleavage, ascot, bare shoulders, thigh strap, white shorts, white vest, gloves, pink jacket, ankle boots, white boots, panther tail</code></pre>any costume<pre><code>raorapanthera, panther girl, panther tail, animal ears, animal ear fluff, blush, pink hair, streaked hair, large breasts, long hair</code></pre><h3 id="limitations-huq4jcbb4">Limitations</h3><ul><li>still cannot remove jacket in debut costume</li><li>clothing details in debut maybe inaccurate</li><li>newyear costume is not working (only twintails is working)</li></ul><h2 id="training-details-54fb674dh">Training Details</h2>LoRA rank<ul><li>reduced dimension to 8 with sv_fro=0.9</li></ul>dataset<ul><li>55 images debut costume</li><li>20 images others</li><li>7 images new year costume (not working)</li></ul>parameters<ul><li>resolution = 1024</li><li>batch size = 2</li><li>dim,alpha = 16,16 (for training)</li><li>mix/save precision = bf16/bf16</li><li>optmizer = AdEMAMix + weight_decay=0.025 betas=0.9,0.99,0.9999</li><li>UNet LR = 2e-4</li><li>TE LR = 1e-4</li><li>scheduler = cosine_with_min_lr (min_lr_ratio 0.33)</li><li>huber snr 0.85</li></ul>steps<ul><li>epochs = 10</li><li>total steps = 1975</li><li>repeat = 3,8,10</li></ul>tools<ul><li>kohya-ss GUI v24.3.0 FORKED by vjumpkung</li><li>torch 2.5.0 cu124</li><li>RTX 3060 12 GB + xformers + gradient_checkpointing</li></ul>avg weight<ul><li>UNet average weights : 0.013589977964081545</li><li>TE1 average weights : 0.011423828994641868</li><li>TE2 average weights : 0.009651259635575112</li></ul>*This LoRA is for studying LoRA training with new technique so do not use for damaging the vtuber (also support her too).*If you see my model reposted on other image generation service please report to me.

Files with SHA256: 14ea8631c68c2ce088c59d98b51f53d9e85479486e3ec05a4c9b105c368a753e