[ AI Art ] - stable diffusion 討論 分享 01

留明

lora weight用好大？

留個名先 今晚再研究

LoRa train 真人 learn rate , unet learning rate 同 text learning rate 分別set幾多，同pre image train 幾多次

整極都唔似樣

用controlNet啦

你list個setting出嚟,sampler? step?,我最近由gtx1650 super 4G上咗3060 12G, test 完話你知

通常 512x512 少過 10 秒

https://www.youtube.com/watch?v=eghGQtQhY38&t=4s

多謝你肯回應。不過就目前而言，唔係太work，多數image to image慢慢都係會走樣，加曬negative prompt都一樣多隻手或者走樣

唔該，有無人識，
點樣令幅圖focus落對波到？

次次出黎對波都成為配角
整咗好耐都做唔到咁：
https://twitter.com/HannahYmm/status/1635313729817706497/photo/1
https://twitter.com/Cuteillust/status/1634375337990828032/photo/1

breasts focus

試過：((focus on breasts)),((breast shot)),boobs pov,Focusing on the chest,Focus on the Bust,Close Up breast ,(((breast focus))),boobs view,

但都做唔到咁嘅效果https://twitter.com/HannahYmm/status/1635313729817706497/photo/1

用
image to image / controlnet 喇

(upper body)試試


如果加左weight都唔得應該要執執d prompt.或者用最簡單既方法controlnet

對波的確有大D：


用controlnet、img2img有時會變形，例如個頭變肥婆：


始終研究唔出，點樣可以好似人地咁，將對波整到就黎飛出黎咁
https://twitter.com/Cuteillust/status/1635762728873893888/photo/1

好想整到對波……應該點形容，就黎跳出黎咁，成為成張圖嘅主角

補link : https://twitter.com/Cuteillust/status/1634678061097709570

高抄from above試下

唔明你問題係點樣唔得
唔係不外乎pov,upper body,from above呢d咩
一係就收窄張相做512x768即upper body

sorry ,係我表達得唔夠清楚。
我意思係，想幅相可以上面link條女咁，
有種爆乳+focus對波嘅感覺，

但一直捉唔到點樣角度或prompt先做到爆乳focus。

可以試下 a girl with (huge breasts:1.2), she is (very close to camera:1.2), (from above:1.2)

其中一D咒文
<lora:Korean-doll-likeness:0.5>,<lora:Japanese-doll-likeness:0.5>,cute,1GIRL,loli:0.3,STudent,,<lora:Taiwan-doll-likeness:0.1>
都用緊0.5

同時用幾個lora依啲weight加埋其實都大weight 好大機會會令到啲圖無咁準 （以我理解係因為啲人train lora嘅時候啲原圖有都衫 所以大weight會影響啲衫）你試下用少啲lora同再少啲weight 例如全部0.1 / 0.2試下會唔會好啲 （唔好理個樣住 純粹試下關唔關lora事）

同埋你可以試下依個少少改良版嘅prompt

a highschool girl in (seifuku:1.2) white uniform, (serafuku:1.3), the uniform is sailor style and has a huge navy blue (sailor collar:1.5), (red bowtie:1.2), (joshi koukousei seifuku), (sailor fuku), pleated skirt

記得CFG value唔可以太細 我自己用9.5

依家試緊將漫畫圖真人化
https://na.cx/i/4FmMQzA.jpg

https://na.cx/i/NGb5mGk.png

all about upscale!


https://phhofm.github.io/upscale/introduction.html