<h1 id="rice-shower-pony-lora-by-vjumpkung-8d624x0ra">Rice Shower Pony LoRA by vjumpkung</h1><ul><li>I found someone release their dataset but it for 1.5 so I gather it and re-classification by myself (without any AI) and train with many concept together and see how it look like. </li><li>For studying LoRA Training only don't offense me.</li></ul><h2 id="changelogs-jo7jsp4q1">Changelogs</h2>V2 : 2e-4 U-Net LR and 1e-4 TE LR AdEMAMix + inverse sqrt scheduler (if you use metadata viewer it shows v4 because I train this for 4 version now but release in civitai only 2 version please don't dm me about v1 and v2 because one is overfitting and another one is underfitting)V1 : Using AdEMAMix8bit optimizer with inverse sqrt scheduler (still underfitting) dataset + caption reference : <a target="_blank" rel="ugc" href="https://civarchive.com/api/download/models/26990?type=Model&amp;format=SafeTensor&amp;size=full&amp;fp=fp16">https://civarchive.com/models/22604</a> and rearrange by me.Recommended options<ul><li>weight : 1</li><li>Trigger words<ul><li>Base (with any costume)<ul><li>rice shower \(umamusume\)</li><li>Headwear and blue rose are not included!</li></ul></li><li>Rice's signature headwear<ul><li>blue headwear, tilted headwear, hat flower, blue rose</li></ul></li><li>Racing competition costume :<ul><li>blue headwear, tilted headwear, hat flower, blue rose, long sleeves, purple sleeves, sleeves past wrists, fur collar, off-shoulder dress, bare shoulders, blue dress, dress bow, sheathed dagger, lace trim, lace-trimmed legwear, brown thighhighs, shoes, black footwear</li><li>(optional 1) steepled fingers</li><li>(optional 2) glowing eye, flaming eye, eye trail</li></ul></li><li>Casual :<ul><li>casual, hairband, hair ribbon, brown ribbon, long sleeves, collared dress, brown dress, frilled dress, frills, white socks, brown footwear, mary janes</li></ul></li><li>Make up Vampire! :<ul><li>official alternate costume, halloween costume, hair flower, blue rose, hair ornament, frilled hairband, purple gloves, striped bowtie, black bowtie, brooch, puffy short sleeves, bat wings, collared shirt, white shirt, center frills, jack-o'-lantern, orange bow, high-waist skirt, medium skirt, purple skirt, spider web print, black skirt, purple footwear</li><li>(optional) claw pose, fangs</li></ul></li><li>Tracen school uniform<ul><li>tracen school uniform, summer uniform, serafuku, puffy short sleeves, bowtie, horseshoe ornament, sailor collar, sailor shirt, frills, white skirt, purple shirt, miniskirt, zettai ryouiki, white thighhighs, brown footwear, shoes</li></ul></li><li>yummy dreamy fairy (V2 Only)<ul><li>yummy dreamy fairy, twintails, long hair, headwear, blush, dress, long sleeves, purple eyes, apron </li></ul></li></ul></li></ul>reference <a target="_blank" rel="ugc" href="https://civarchive.com/models/22604">https://civarchive.com/models/22604</a><h2 id="training-details-(v2-only)-fqt4p1jdf">Training Details (V2 only)</h2>LoRA size<ul><li>dynamic dim/alpha (maximum 8)</li></ul>dataset<ul><li>Racing competition costume : 150 images</li><li>Make up Vampire! : 50 images</li><li>Casual : 30 images</li><li>Tracen School Uniform : 27 images</li><li>Yummy Dreamy Fairy : 20 images</li><li>Other : 62 images</li></ul>*add repeats until close to 150 images for fixing imbalance datasetparameters<ul><li>resolution = 1024</li><li>batch size = 2</li><li>dim,alpha = 16,16 (for training)</li><li>mix/save precision = bf16/bf16 -&gt; fp16 for reducing</li><li>optmizer = AdEMAMix (Mega Consume VRAM from 10.5 GB to 11.5 GB)</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 = 4380 steps</li><li>repeats = 1,3,5,6,7,2 (following dataset)</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: 0.015112816238817505</li><li>Text Encoder (1) weight average strength: 0.011408020326133637</li><li>Text Encoder (2) weight average strength: 0.010726570189826059</li></ul>*This LoRA is for studying LoRA training with new technique so do not use for damaging the character (Do not generate any NSFW).

[Pony] Rice Shower (umamusume)

Learning Rate (LR): UNet LR: 0.0002 TE LR: 0.0001Optimizer: bitsandbytes.optim.ademamix.AdEMAMixScheduler: inverse_sqrt Warmup steps: 100

Files with SHA256: 884529f60a1fde9ed66a71b6ec843cc1264ea36ca486e1b4bd38d211b8b36b29