<h3 id="heading-679">For previous version</h3><p>V1 trained with 6400*3+6400*2 steps, network rank is 128,no flipped img in the dataset.</p><p>V2(my preference for now) trained with 6000*2+6000*2 steps, network rank is 128, flipped img in the dataset.</p><p>V2-lite is a lite version of v2,6000*2 steps. how lite actually？Just lose 93.75% of the "weight".</p><h3 id="heading-680">New version : v3n</h3><p>Set network alpha much larger, which largely effect learning rate.</p><p>Outputs result better and more stable , which could means less stylized.</p><p><span style="color:rgb(194, 189, 181)">Also perform well with simple prompts (last 1 post for example)</span></p><h3 id="heading-684"></h3><p><strong>Without trigger words the outputs result slightly different, maybe worthwhile to try.</strong></p><p>Perhaps you noticed that the negative prompt I used in the posts seems needless,</p><p>well,that's because it's designed for AOM3(nsfw yes)</p><p>By the way,you can easily call out a big cat to accompany her.</p>

Shushou 珠晶

<p>trained by mistaken, actually not that bad and really light-weight</p>

Files with SHA256: b9164dc642d6a1c2966ea260f27fc40405891c390e2e4b4d0d27ef34f599a70d