I'm training dreambooth model with SD-768.
Results seems to overcooked too fast.

75 refs images
1000 regs
3000 steps

Prompt: "portrait of santa claus, arcane style"

0 4

QingShe (Green Skake) in White Snake 2. Dreambooth (left) vs original(right).

1 3

some more game characters
I should test on some point reusing them in dreambooth to further finetune them eg for faces
Could be a nice and nearly fully automatic pipeline for in

0 1

dreamboothで追加学習させる時のプロンプトの仕組みどうも勘違いしてたっぽい
少し仕組み勉強し直してこの前と同じ画像でアメヤさん学習し直したらかなりそれっぽいの出てくるようになった
セーラー服っぽいのも着せれたしパジャマ姿にもできたとはいえ元画像への引っ張られ方が前よりきつい感ある

0 2

ダイヤさん
超解像のみ

3 16

new grid who dis

prompt is just "[trained tokens] portrait" 🤯 that's almost dreambooth quality

0 15

AI…頼む…甜甘を…描いてくれ…!

22 98

8528diffusion のv0.2を作る過程で、別モデル学習の検証。WDへの学習データには自分が描いた絵(二次創作含む)を利用。#8528d より学習データ枚数が足りないのでめちゃくちゃ傾向が偏るのと制御難易度跳ね上がり。 "自分の絵で簡易追加学習する"ならdreamboothの方がいいかもな……

8 138

experiments. The training set was a bunch of random SD images I made and liked.

0 1

experiments. The training set was a bunch of random SD images I made and liked.

0 2

Another experiment. I guess I shouldn't say "sexy female android" if I don't want nudity.

0 0