Need Help, Look at the Embedding I trained.

[LanceGao97]

I spent 8 hours study embedding, followed tutorials, and read articles , this is so shit result.

I just imput 32 PNG pixelart nice looking portrait from a game called KingOf fighters2000, Hope to train a pixel art style ot be used on every portrait,
However, the result, the only thing comfort me is that it does tranlate everything to pixel style art,
but its so unreasonable, I don't know how to control , every parameter, so many parameters , none of them is understandable.

Can anyone give me a hint ?

the part im confused is the

initial text
the frightening default step number, 10,0000, I trained it 3000 steps, most people from tutorials say 3000 is enough.
It does can finish， but I don't what the hell it is doing.
I simple just wanna train a style of pixel art for portrait, it shouldn't be that fancy or hard for this StableDiffusion， since all those realistic things can be trained already.

[LanceGao97]

the input images are like this,
the embedding template is a .txt,
a pixel style portrait of [filewords], art by KOFF3 style,
the filewords are then replaces by the tags in each image, I can see the process on webui.

[elen07zz]

mm try using lora instead , maybe that could work better.
https://civitai.com/?query=pixel&view=feed

[LanceGao97]

mm try using lora instead , maybe that could work better. https://civitai.com/?query=pixel&view=feed

OK， i still think the way textual inversion is doiing is so strange and unpredictable, but Easynegative is an embedding, and Easynegative is so powerful, I don't know how it can do so.