<h1 id="yuuki-sakuna-()-pony-lora-v1kqs1zvy">Yuuki Sakuna (結城さくな) - Pony LoRA</h1><h2 id="disclaimer-1pcixkk28">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><h3 id="version-3-description-sd947q9wa">Version 3 description</h3><ul><li>reduce dim/alpha to 16/12</li><li>using scale weight norm to 1.5 (prevent overfitting)</li><li>change scheduler</li><li>new dataset + captions</li><li>her smile is less become :3 instead of :)</li></ul><h3 id="trigger-word-(v3-only)-ea7s63qut">Trigger Word (V3 only)</h3><ul><li>default costume</li></ul><pre><code>yuuki sakuna, long hair, blush, animal ears, pink hair, medium breasts, hair ornament, maid headdress, bow, maid, animal ear fluff, puffy sleeves, apron, white dress, bowtie, collarbone, hairclip, hair bow, detached collar, black footwear, shoes, black choker, white thighhighs</code></pre><ul><li>any costume</li></ul><pre><code>yuuki sakuna, long hair, blush, cat ears, pink hair, medium breasts, hair ornament, bowtie, hairclip, hair bow</code></pre><h3 id="limitations-pipkivzow">Limitations</h3><ul><li>some costume cannot change properly (due to lack of datasets)</li><li>full body still inaccurate (but much improvement)</li></ul><h3 id="version-2-description-6i3pe6lug"> Version 2 description</h3><ul><li>using Experimental Optimizer (it seems improvement on full body but camera angle still not variable if not prompt it)</li><li>bow still remains even change outfit</li></ul><h3 id="version-1-description-zbreg0d1p">Version 1 description</h3><ul><li>I like cat ears girl so, I trained as LoRA (don't know about her past at all :&lt;)</li></ul><h3 id="trigger-word-fx8j8trfj">Trigger Word</h3><ul><li>trigger word</li></ul><pre><code>yuuki sakuna</code></pre><ul><li>any costume (still not flexible enough)</li></ul><pre><code>yuuki sakuna, long hair, animal ears, pink hair, blush, cat ears, pink eyes, two side up, ahoge, colored inner hair, two-tone hair</code></pre><ul><li>debut costume (good but some component still missing)</li></ul><pre><code>yuuki sakuna, long hair, hair ornament, bow, animal ears, pink hair, blush, cat ears, maid headdress, hair bow, frills, hairclip, pink eyes, pink bow, blue bow, maid, puffy sleeves, two side up, cat hair ornament, ahoge, heart hair ornament, puffy short sleeves, clothing cutout, pink dress, blue bow, colored inner hair, two-tone hair, cleavage, breasts</code></pre><ul><li>for full body add this following text will help (shoes still not correct)</li></ul><pre><code>shoes, black footwear, white thighhighs</code></pre><h3 id="limitations-x8amc6eyy">Limitations</h3><ul><li>Cannot change costume properly (still have some debut costume component leftover)</li><li>full body may be not effective</li><li>LoRA still little bit underfitting (like medium rare pork) (in version 1)</li><li>version 2 is improve some small detail but dataset still not variable enough (due to imbalance image)</li></ul><h3 id="training-details-(version-3)-yhlrkivgg">Training Details (Version 3)</h3>LoRA size<ul><li>reduced dimension to 8 with sv_fro=0.95</li></ul>dataset<ul><li>42 images</li></ul>parameters<ul><li>resolution = 1024</li><li>batch size = 2</li><li>dim,alpha = 16,12 (for training)</li><li>mix/save precision = bf16/bf16</li><li>optmizer = AdEMAMix + weight_decay=0.025 betas=0.9,0.999,0.9999</li><li>UNet LR = 2e-4</li><li>TE LR = 1e-4</li><li>scheduler = cosine_with_min_lr (min_lr_ratio 0.67)</li><li>huber snr 0.85</li></ul>steps<ul><li>epochs = 5</li><li>total steps = 1575</li><li>repeat = 15 (one concept only)</li></ul>tools<ul><li>kohya-ss GUI v24.3.0 (Forked by me)</li><li>torch 2.5.0 cu124</li><li>RTX 3060 12 GB + xformers + gradient_checkpointing</li></ul>weight<ul><li>UNet average weights : 0.0149531283161857</li><li>TE1 average weights : 0.011002991641968642</li><li>TE2 average weights : 0.009832777519477531</li></ul><h3 id="training-details-(version-2)-4v9gu04bq">Training Details (Version 2)</h3>LoRA size<ul><li>reduced dimension to 8 with dynamic alpha</li></ul>dataset<ul><li>38 images (most is half body)</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 (32 bit consume VRAM)</li><li>UNet LR = 2e-4</li><li>TE LR = 1e-4</li><li>scheduler = inverse_sqrt with warmup 100 steps</li><li>l2 loss only</li></ul>steps<ul><li>epochs = 10</li><li>total steps = 2850</li><li>repeat = 15 (one concept only)</li></ul>tools<ul><li>kohya-ss GUI v24.2.0</li><li>torch 2.5.0 cu124</li><li>RTX 3060 12 GB + xformers + gradient_checkpointing</li></ul>weight<ul><li>UNet weight average strength = 0.015634962041489377</li><li>Text Encoder (1) weight average strength Clip_L = 0.011193290141749815</li><li>Text Encoder (2) weight average strength Clip_G = 0.010691167576002698</li></ul><h3 id="training-details-(version-1)-33fq1i30y">Training Details (Version 1)</h3>dataset<ul><li>38 images (most is half body)</li></ul>parameters<ul><li>resolution = 1024</li><li>batch size = 2</li><li>dim,alpha = 16,16 (no resize for preserving quality if LoRA is good enough will do it :P)</li><li>mix/save precision = bf16/fp16 (accidentally change)</li><li>optmizer = AdEMAMix8bit</li><li>UNet LR = 1e-4</li><li>TE LR = 5e-05</li><li>scheduler = inverse_sqrt with warmup 100 steps</li><li>huber snr with c = 0.85</li></ul>steps<ul><li>epochs = 10</li><li>total steps = 2850</li><li>repeat = 15 (one concept only)</li><li>full_bf16 training</li></ul>tools<ul><li>kohya-ss GUI v24.2.0</li><li>torch 2.5.0 cu124</li><li>RTX 3060 12 GB + xformers + gradient_checkpointing</li></ul>weight<ul><li>UNet weight average strength = 0.008335085569112463</li><li>Text Encoder (1) weight average strength Clip_L = 0.0073367764333498705</li><li>Text Encoder (2) weight average strength Clip_G = 0.005826970830639767</li></ul>description ref from <a target="_blank" rel="ugc" href="https://civarchive.com/user/Gtonero">Gtonero</a>*This LoRA is for studying LoRA training with new technique so do not use for damaging the vtuber (also support her too).

[Pony] Yuuki Sakuna (結城さくな)

<ul><li>Update some LoRA training parameters</li></ul>

Files with SHA256: ee0d8f58d2ad602fd8bbc1440be37696f1371d1c3964948913d1dc0517fd4dc3