<h3 id="heading-623">The updated model must be used on used on webui version &gt;= 1.7.0</h3>Intermediate checkpoints <a target="_blank" rel="ugc" href="https://huggingface.co/alea31415/random_stuff/tree/main/tearmoon">https://huggingface.co/alea31415/random_stuff/tree/main/tearmoon</a>This model uses my updated pipeline <a target="_blank" rel="ugc" href="https://civarchive.com/articles/2383">https://civarchive.com/articles/2383</a> which constructs the dataset automatically from anime.---Wildcard:<pre><code>Mia
Mia, frills, blue dress, puffy short sleeves, blue bow
Mia, necktie, school uniform, collared shirt, pleated skirt
Mia long hair
Mia long hair, blue dress
AnneT
AnneT, maid headdress, maid apron
SionT
SionT,epaulettes,white gloves
ChloeT
ChloeT, school uniform, glasses, red necktie, collared shirt, white shirt
Tiona
Tiona, dress
Keithwood
Keithwood, black cape
Elise
Elise, glasses, yellow shirt, collarbone
Ludwig
Ludwig, glasses, ascot
AbelT
AbelT, epaulettes
GainT
GainT, military uniform, holding sword
Rafina
Rafina, white dress
DionT
DionT, armor, epaulettes
DionT helmet
DionT helmet, armor, cape, knight
Liora
Liora, maid apron, black dress, juliet sleeves, hair ornament, leaf hair ornament
Cyril
Cyril, blush, suspenders
Lynsha
Lynsha, juliet sleeves, hair bow
Lambert
Lambert, red bowtie
screenshots
booru</code></pre>No balancing at character appearance level cause the failure in learning secondary appearance. For <code>mia long hair</code> it is better to use later checkpoints from hugging face.This is another experimental run for OFT where I set unet lr to 1e-4, text encoder lr to 2e-5, and embedding lr to 1e-2. The higher learning rate might hurt flexibility.

Files with SHA256: 1ff1d1cf5a36da529785e3c1289b6ebac77162d9910855e6393ba658ab748837