flat anime slider

2020 japanese anime slider

Still using Slider Lora

Extracting animations from Z-images is bloody difficult

Actually, it's best not to set the weight too high—0.4 to 0.6 is perfectly sufficient. Increasing it further can actually cause distortion, at which point you can utilise trigger words.

I've used tags from images on the Danbooru website, which should do the trick.

一樣是slider lora

要從Z-image引出動畫有夠難的

其實權重最好不要太高0.4-0.6就很夠了,再加下去反而會破壞,這時候就可以用triger word

我用過danbooru網站上面圖的TAG應該是可以了,

---

---

First test utilised a slider approach for training within Z-image.

Quite straightforward.

Positive prompt: Flat-painted animated illustration.

Negative prompt: Hyper-realistic photograph.

That's all.

Is it effective? Yes, when weights are set to 1, provided the prompt contains no redundant terms like 'painting'/'animation'/'photography' etc.

For added assurance, one could incorporate prompts such as 'animated character portrait'.

However, I've noticed something: Z-image prompts exert an extremely potent influence, potentially overriding LORA's effects.

Increasing LORA's weight rapidly amplifies its impact. Around 1.1, the difference becomes pronounced. Beyond 1.5, the prompts struggle to overpower LORA, yet the results become rather peculiar.

Therefore, the prudent approach is to set the weight between 1 and 0.9, while adding prompts like "illustration" or "anime".

Then, how is it that the Chinese prompt yields better results?

第一次測試在z-image之中使用slider方式進行訓練

很簡單

正面概念：平塗動畫插圖

負面概念：超寫實照片

就這樣

有用嗎？有用，權重設到1，如果提示詞中沒有任何多餘的詞彙如繪畫／動畫/攝影照片等

如果要保險一點當然可以加上動畫人物肖像之類的提示詞

但是發現一件事情Z-image的提示詞效果非常的強烈,有可能會蓋過lora的效果

如果提升lora權重,lora影響力飛速上升大概到1.1左右就很明顯1.5以後提示詞會壓不過LORA但是會很怪異

所以保險是權重1-0.9 提示詞加一個illustration或者anime之類

然後,中文提示詞效果更好是怎樣