LoRA of Narita Taishin (umamusume)All of my works is free. If you'd like to support me, feel free to <a target="_blank" rel="ugc" href="https://www.buymeacoffee.com/mhtLoRA">buy me a coffee</a>. 👍Recommend options<ul><li>LoRA weight 0.7~1.0</li><li>Trigger words<ul><li>base<ul><li>narita taishin \(umamusume\)</li><li>(optional)closed mouth, serious<ul><li>These tags make her kawaii.</li></ul></li></ul></li><li>Racing competition clothes :<ul><li>torn jeans, (asymmetrical clothes:1.1), belt, fur-trimmed jacket, jacket around waist, clothes around waist, midriff, long sleeves, tied shirt, open clothes, yellow shirt, mismatched footwear, asymmetrical footwear<ul><li>asymmetrical tags might be need higher weight. (It is challenging for AI)</li></ul></li></ul></li><li>Casual clothes :<ul><li>black headwear, white shirt, red jacket, long sleeves, open clothes, torn pants, denim, blue pants</li></ul></li><li>Steampunk costume :<ul><li>white shirt, long sleeves, center frills, frilled shirt, brown belt, corset, monocle, semi-rimless eyewear, fingerless gloves, official alternate costume, steampunk</li></ul></li></ul></li><li>Use animefull based model</li><li>Settings<ul><li>DPM++ Series(SDE Karras, 2M Karras, etc.)</li><li>about 20 steps, CFG scale 3.5~6.5</li><li>kl-f8-anime2.vae</li><li>CLIP skip = 1</li><li>If using highres<ul><li>Latent (bicubic antialiased)</li><li>Denoising strength 0.55~0.7</li></ul></li></ul></li></ul>LoRA training info<ul><li>based on Animefull-pruned</li><li>108 images of narita taishin<ul><li>8x low quality 32 images</li><li>16x medium quality 32 images</li><li>32x high quality 44 images</li><li>=&gt; 1 epoch = 2176 images</li></ul></li><li>replaced character feature tags with narita taishin \(umamusume\)<ul><li>horse girl, horse tail, brown hair, blue eyes, etc.</li></ul></li><li>using aspect ratio bucketing</li><li>resolution : 768x768</li><li>rank=alpha=16</li><li>using <a target="_blank" rel="ugc" href="https://github.com/facebookresearch/dadaptation">dadaptation</a></li><li>batch size = 4, 9 epochs trained</li><li>clip skip = 1</li></ul>This LoRA may be compatible with animefull-based models (ex. AbyssOrangeMix Series...)All uploaded images are output of AbyssOrangeMix2-nsfw, kl-f8-anime2.vae

Files with SHA256: 7cb610b1b74391f52c40785c305e4eec573ff384d64406ac460ec876efb87e92