小日向ちゃんっぽいガールと同時に生成される2P小日向ちゃん

0 2

ふんいき似てるけどサスペンダーを理解してないAIイラスト

0 3

Added another embedding, anthro which has a totally different effect from when using it in a simple way. In this case it adds shading.

not

0 3

より小日向ちゃんっぽいのが生成された。画像サイズ512×768良い?

9 36

TextualInversionで学習したものを使用して描画。
リアル系だとあまりうまくいかないのでアニメ調で描画してみる😆

0 5

wd1.4 + textual inversionで運良く生成できた和服響子ちゃん
サイドテールの位置、横髪の長さ、前髪の分け目、とかが左右非対称でTIだと潜在空間でバッチリ指定できる概念じゃないんだろうな

0 0

細部が詰めきれてないけどwaifu-diffusion-1.4-anime_e1でtextual inversionがまずまず成功した。
(見た目に関するプロンプトはTIに押し込めた)

0 3

From 90 to 126k it still produces poor results, but it may be workable. I'll try retraining anyway, I guess. REQUIRES a CFG scale lower than 10 and calling the embedding near the end of the prompt to be adequate.
https://t.co/3n45LipS0E
https://t.co/b8Jh5SPH5U

2 2

Despite a very consistent style, the results still fail to match it most of the time at 78k. Besides hair, it also cannot discern the requested color of the vest, which is rather alarming. I may have to retrain it from scratch.
https://t.co/FJbHkWKYbZ

0 0

66k steps with last 10k done at 1.0e-04 learning speed, getting closer and more consistent now.
https://t.co/UEs1bQMVBJ

0 1

asutora_gs31800 -> no embedding -> miata8674_gs45000 -> atawashi_gs56000
https://t.co/GDA22hOKdt

0 2

By 56k steps, it still hasn't learned much but the coloring method, and the results are mostly a mess.
https://t.co/0g3ZlCfJKW

0 0

Training continues, 33k steps now. For some reason it keeps trying to give her yellow hair. The artifacts are severe. Still quite a ways till it's passable even at CFG 8. Fourth pic is with the style invoked last, the rest are early in the prompt.
https://t.co/zDB4ifodHV

2 0

Kiriko from Overwatch 2
1.3 embedding
>100 source images, 50k steps
just paste the file into the embeddings folder of the webui and prompt with "kiriko"
https://t.co/FUSoSs518K

0 5

Same parameters but different styles invoked in the prompt, clockwise from top left:
No embedding —> miata8674_gs45000 ⮒
asutora_gs31800 —> (training-in-progress) atawashi_gs22000.
https://t.co/ohCrjIRRbL

1 0

I soon realized I was calling it wrong, without using the underscore. Nevertheless, I find it's still worse than my default negatives.
https://t.co/Jf2340JfDb

1 1