[ AI Art ] - stable diffusion 討論 分享 02

1.5*

主體係oldfish
唔記得有冇溝其他model

即係我係咪要放d 有角嘅相落去train, 咁喺其他相，例如貓貓狗狗嘅相就會得出有角嘅結果

所以你練果時要俾埋tag 佢，等佢唔好咁易train 到所有野都有角。

Civital 上面成日有啲LoRA 亂鳩咁落tag又overtrained ， 好易會搞到貓貓狗狗都變埋獨角獸。

用下mov2mov 整動畫, 唔落ebsynth似乎都係好難解決flickening 問題...用multiple control net 又有bugs....

have u seen my test?
Ebsynth is much better than mov2mov

想問下見到D AI KOL佢地可以整到個臉張張差唔多樣係點整出黎
定係佢地都係用明星既lora合成整到差不多樣

ebsynth 用過,只係懶D 想係webui 一野過輸出 , 純粹試下, 見D大陸人玩mov2mov 玩得好勁....

This is mov2mov, but not as bad as yours.

其實唔難

未出Ebsynth Utility果陣玩過一陣放棄左, 依家有懶人包試下先.....

未試Ebsynth Utility,我用嗰個係另外一個.
Ebsynth有個致命問題係每次keyframe最多20張,要人手分batch做, 同埋佢個ui好廢

等緊佢出Ebsynth Studio 1.0

咩setting可以咁consistent

multicontrolnet + denoise唔好太高
controlnet我用咗 canny + depth

咁要爬下YouTube
我明明地，UI 唔friendly 唔記得點整

no prompts?

this is my prompts

korean girl

with some basic negative prompts....

Denoise set得唔高其實冇咗mov to anime嘅意義。

見啲日本仔gen出黎幾好睇喎， 試吓先

理論上freeze (masked)左個background , 落兩個control net ,denoise 可set 高啲，再用ebsynth blend 理一齊應該真係做到animation 效果，不過未試到，local 機得6GB vRam 整幾野就爆

yes, but judge by result, not the denoise value,
set the denoise value range I think 0.2-0.5
to a point that the image is change to the style you want of course.

有冇D prompts 可以減少係咁閃，件衫都係咁轉