[SD 1.5][LORA]Curren Chan (umamusume) V1.0
<p>LoRA of Curren Chan (umamusume)</p><p></p><p><strong>NOT COMPATIBLE WITH AOM3A2, AOM3A3 model.</strong></p><p></p><p>Recommend options</p><ul><li><p>LoRA weight 0.6~1.0</p></li><li><p>Trigger words</p><ul><li><p>base</p><ul><li><p>curren chan \(umamusume\)</p></li><li><p>(optional)smile, hands on own chin,</p></li></ul></li><li><p>Racing competition clothes :</p><ul><li><p>vertical-striped dress, bare shoulders, off shoulder, closed mouth, long sleeves, collared dress, gloves</p></li></ul></li><li><p>Wedding costume :</p><ul><li><p>elbow gloves, black wedding dress, black gloves, string of fate, official alternate costume, closed mouth, bare shoulders, sleeveless dress</p></li></ul></li><li><p>Casual clothes :</p><ul><li><p>jewelry, white shirt, white long sleeves, necklace, red skirt, sleeves past wrists, black pantyhose</p><ul><li><p> maybe white shirt need more weight</p></li></ul></li></ul></li></ul></li><li><p>Use animefull based model</p></li><li><p>Settings</p><ul><li><p>DPM++ Series(SDE Karras, 2M Karras, etc.)</p></li><li><p>about 20 steps, CFG scale 3.5~6.5</p></li><li><p>kl-f8-anime2.vae</p></li><li><p>CLIP skip = 1</p></li><li><p>If using highres</p><ul><li><p>Latent series</p></li><li><p>Denoising strength 0.50~0.65</p></li></ul></li></ul></li></ul><p></p><p>LoRA training info</p><ul><li><p>based on Animefull-pruned</p></li><li><p>133 images of curren chan</p><ul><li><p>4x low quality 42 images</p></li><li><p>8x medium quality 48 images</p></li><li><p>16x high quality 43 images</p></li><li><p>=> 1 epoch = 1240 images</p></li></ul></li><li><p>replaced character feature tags with curren chan \(umamusume\)</p><ul><li><p>horse girl, horse tail, pink hair, purple eyes, etc.</p></li></ul></li><li><p>using aspect ratio bucketing</p></li><li><p>resolution : 768x768</p></li><li><p>rank=alpha=16</p></li><li><p>using <a target="_blank" rel="ugc" href="https://github.com/facebookresearch/dadaptation">dadaptation</a></p></li><li><p>batch size = 4, 9 epochs trained</p></li><li><p>clip skip = 1</p></li></ul><p></p><p>This LoRA may be compatible with animefull-based models (ex. AbyssOrangeMix Series...)</p><p></p><p>This LoRA can generate nsfw images of curren chan.</p><p></p><p>All uploaded images are generated by AOM3A1</p>
页:
[1]