[ AI Art ] - stable diffusion 討論 分享 01

表情關鍵詞

https://note.com/misimo_ai/n/n359f9bd7d0df?fbclid=IwAR32UsQq29GvE6hqJuhnt7OVgoCdaAPxzLUc-tkS5wtGYrhpbyhWd0EAkok

I tested with this >
lower body in water


bottom half in water

唔該巴打
有冇腰以下浸喺水個咒語

呢張?

(bottom half in water:1.5),<(realistic:1.4), (extremely intricate:1.2)>, 1girl, solo, grey background, hair ornament, school uniform, short hair, sleeves past fingers, pleated skirt, shirt, hairclip, bangs, black skirt, looking at viewer, white shirt, closed mouth, brown eyes, brown hair, grey hair, black sailor collar, cowboy shot, black necktie,
Negative prompt: (worst quality, low quality:1.4), (depth of field, blurry:1.2), (greyscale, monochrome:1.1), 3D face, cropped, lowres, text, jpeg artifacts, signature, watermark, username, blurry, artist name, trademark, watermark, title, multiple view, Reference sheet, curvy, plump, fat, muscular female, strabismus, clothing cutout, side slit, (nsfw:1.3),
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 9, Seed: 130604522, Size: 512x768, Model hash: aff40917d5, Model: ultraILLUSI0N2D_ultraILLUSI0N2DV1, Clip skip: 2, ENSD: 31337

civitai終於有得hide anime content

玩下呢個 ControlNet 姿勢圖片庫,好用
https://civitai.com/models/18075/1000-boudoir-glam-shot-poses

原來係咁
我終於明白
我一直以為係要用img2img入面既inpaint油黑要轉既部份去轉真人
多謝兩位巴打暢教
試左一大輪之後確實發現 denoise 0.35~0.55 炒車率係最低

岩岩試左
係幾好幾sharp
但係同latent比真係差咁D

就好似全張相都用左LLuL咁

係咪用唔到controlnet??
用左之後個姿勢圖係全黑色

都係差唔多咁做法，不過我有再mix過吓啲model咁，等佢兩個model concept 唔會差太遠。

動漫轉真人原因係通常動漫風嘅通動作，衣著等等都比真實風有創意好多，可以做到啲獨特啲嘅圖。

但缺點係人物比例會好易出錯

自己mix咗三個幾滿意嘅model
偏真實:

偏2.5D:

下面依個有少少露點
偏厚塗:
https://na.cx/i/KLwZgTF.png
大家鍾意邊個多啲?

用咩prompt 出到咁真人樣?

Lora 嘅話15張100次就夠minimum requirements
理論值係1500/images number 因為本身都幾overfit 下

冇 最多睇size checkpoint 係以GB計 Lora 係以MB 計

用咩prompt 出到咁真人樣?

裝 dreambooth 話我 xformers 太舊
唔知點搞

你夠ram 玩dreambooth?

裝得xformer 似ram少先用

其實重點在於你本身用緊chilloutmix model
系統會默認出圖係真人 而唔係2d人像
咁你係 img2img 是但打d 咩 young japanese girl, long hair, fairer white skin, smiling 依d佢已經會幫你由2d轉真人

俄妹+日本妹+韓妹 LoRA 混合

依個lora幾搞嘢
Xiaorenshu_v20

轉一次去半寫實既好似會冇咁怪

我而家試緊啲好瘋狂嘅野, 直接用img2img,唔洗訓練Lora 就可以regenerate 新嘅pose但keep八成至九成個樣同服裝.
再確認同測試下....

想問下點解我google stable diffusion 搵到嘅網頁都用唔到嘅
有冇人可以俾link 我