search

p1atdev / LECO

Low-rank adaptation for Erasing COncepts from diffusion models.
https://arxiv.org/abs/2303.07345
Apache License 2.0
307 stars 23 forks source link

unconditional and neutral #23

Closed PotatoBananaApple closed 1 year ago

PotatoBananaApple commented 1 year ago

Hey,

Would it be possible to have example how would one use unconditional and neutral?

word to take the difference from the positive concept

starting point for conditioning the target

These don't really open it for me sadly. 😥

sdbds commented 1 year ago

target = trigger words you will use such as lora

positive = postive prompt unconditional =negativae prompt

train images = (positive - unconditional)* cfg + neutral

example: i want to make a anime girl image to replace 1girl positive prompt :1girl, (masterpiece:1.2) negative prompt : (3d:1.2)

so target =1girl positive = masterpiece negative = real neutral = 1girl CFG = 1.2

PotatoBananaApple commented 1 year ago

Thank you for answer.

I was reading the projects linked in reference. In the concept mod there was mention about the unconditional and it read like this

Write to Unconditional: To write a concept to the unconditional model, use the "=" operator after the concept. Example: "alpaca=" causes the system to treat "alpaca" as a default concept or a concept that should always be considered during content generation. -https://github.com/ntc-ai/conceptmod

But what you said makes more sense to the word to take the difference from the positive concept