[ AI Art ] - stable diffusion 討論 分享 02

未玩過train 睇上面巴打話用dreambooth咪試下

岩岩試
(8k, best quality, masterpiece:1.2), (realistic, photo-realistic:1.37), ultra-detailed, 1 girl, cute, solo, beautiful detailed sky, detailed cafe, night, sitting, dating, (nose blush), (smile:1.15), (closed mouth) small breasts, beautiful detailed eyes, (collared shirt:1.1), business attire, (short hair:1.2), floating hair NovaFrogStyle, <lora:taiwanDollLikeness_v10:0.2>, <lora:koreanDollLikeness_v15:0.2>, <lora:japaneseDollLikeness_v15:0.2>, <lora:breastinclassBetter_v141:0.3>,

Negative prompt: EasyNegative, paintings, sketches, (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, manboobs, backlight,(ugly:1.331), (duplicate:1.331), (morbid:1.21), (mutilated:1.21), (tranny:1.331), mutated hands, (poorly drawn hands:1.331), blurry, (bad anatomy:1.21), (bad proportions:1.331), extra limbs, (disfigured:1.331), (more than 2 nipples:1.331), (missing arms:1.331), (extra legs:1.331), (fused fingers:1.61051), (too many fingers:1.61051), (unclear eyes:1.331), bad hands, missing fingers, extra digit, (futa:1.1), bad body, NG_DeepNegative_V1_75T,pubic hair, glans, (outdoor:1.6),
Steps: 25, Sampler: Euler a, CFG scale: 7, Seed: 3896049075, Size: 512x512, Model hash: 26d8b87829, Model: beautypromix_v1



Used embeddings: easynegative [119b]

Time taken: 2m 21.65s



無display card都有得玩, 夠我新手試玩....
覺得岩玩可以再諗換一set新電腦

肯定你會換
我係gtx1650 super 4gb,買咗3060 12gb
我都估唔到買卡唔係為咗打機

其實係要兩樣一齊做
Eg 如果你最終目的係想你女神大解放
你要用dreambooth train 個 checkpoint
再可以train Lora ，我就Lora weighting 0.2 咁加入去，似好多架

你玩gen 一個唔識既人，要求可以低好多
當你要train textual investing , dreambooth，就幾食VRAM

用lora train 其實efficient 好多，佢都係簡化版dreambooth , 容量細好多，佢係base on 基礎model (e.g chill out mix) 加你source 既相一層疊一層train 個network 出嚟, 佢每層都學習緊，而佢train 出嚟只係一個Lora network , 唔係output 另一個基礎model , 如果你只係train 緊1個女神圖象，而唔係1堆人或者一款特定style , 理論上Lora efficient D.

Dreambooth 直頭係修改埋原有基礎model 既數據output 一個新既model , 理論上係會像真d , 佢train 出嚟既新model 係會base on 你既input ，不過要16GB vRam 先train 到，run 既時候長d, 個file 下下幾GB,用落又未必夠Lora flexible

Dreambooth 我係咁出 xformers is not available
研究下lora先

之前玩過吓dreambooth， 直頭係對hard disk 嘅考驗

laptop勸你唔好諗training

gaming notebook 玩到嘅, 但lora出到嚟唔係太似

玩左一個星期
終於 Train 左第一隻 Lora...
好滾動....

伸手黨 有冇一啲常用model/lora/negative prompt

置頂連結 置頂連結

some common <textual inversion> for negative prompts


https://civitai.com/models/7808
https://civitai.com/models/16993
https://civitai.com/models/4629
https://civitai.com/models/4629

model / lora 太多你自己搵 https://civitai.com/, most download.
negative prompts #114 https://lih.kg/zHjFsAX

Negative textual inversion 我聽到時都哇左聲，好有用

[ AI 加持 - mov2mov , ebsynth test ]
近排都係keep住玩movies, 今次用ebsnyth utility幾方便

下面係 mov2mov / ebysnth utility



mov2mov:
撳個掣就搞掂,淨係等,閃到七彩
https://www.dropbox.com/s/whr2jhqnaukrvwf/yujin_mov2mov.mp4?dl=0

ebsynth:
過程繁複, 每個步驟都要人手做, 畫面好順但係有啲鬼影
https://www.dropbox.com/s/0u7533s0gdhll9x/yujin-ebsynth.mp4?dl=0
部分keyframes

郁動冇咁多，衫冇太多pattern 嘅source 易好多整

等我挑戰mtv

masterpiece, best quality, realistic, photorealistic,ultra detailed,
solo,1girl, luxury palace background, irene1, <lora:irene_V70:1> , middle breasts, pink nipples, <lora:lightAndShadow_v10:0.7>, pink pussy, <lora:femaleMasturbationBoob_v1:1>, a female masturbating, pussy, vagina, girl masturbating, <lora:mLegsPoseLora800_v10:1>, photon mapping, radiosity, one girl masturbating, automatic white balance
moist lips, naked, nsfw,missionary,((spread legs))((masturbation)), ((spread pussy by hands)) ,((masturbation)), perfect woman, perfect girl face, <lora:realisticVaginasGod_godpussy1V2:0.3>


(worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome)), mastoptosis, pendulous breasts, bad-hands-5, bad-picture-chill-75v, extra pussy, extra belly button, pubic hair, deformed pussy, bad pussy, clothes, pants, underwear

求教2張圖係同一次gen
圖1係load到92%（好鮮色
圖2係load到100% 中間發生咗乜事呀
色調突然由鮮色好靚變到爛鬼文藝feel
樣又變到衰咗