Training settings - What should they be for a good negative embedding?
Hi. I'm trying to make a negative TI of my own for a few specific issues I want to solve with a lora of mine, among a couple other problems I find in various other models and Loras.
I found this embed through the negative embedding image dataset from your account here: https://huggingface.co/datasets/Duskfallcrew/Negative_Images
Most of those images seem the haven't been processed for training yet and so are lacking captions and that sort of thing. According to huggingface this negative embed used those images for training, so I am curious what steps you made for processing the images.
Furthermore, what settings were used for training is something I can't quite figure out on my own.
Would you be willing to share some insights into your process for me? I just can't seem to get a good embedding that does what it's supposed to, for the most part, so I would certainly appreciate your thoughts (and maybe a portion of the processed images with their caption files if that isn't too much to ask)
OH, LMAO my negative embed sucks lmao - I've never made one that works - you're welcome to use the dataset, I do'nt understand how TI's work that well so i'm tryin to develop neg Loras instead XD
Oh, ok. Literally nobody I have asked has bothered responding to any comments I have made anywhere. Seems like making a decent negative embed is an industry secret or something. Thanks anyway though
No it's more or less that you're looking at knowing what data you want in it, understanding how the Text inversion in A1111 works, crying when ti takes too long - realizing you can just -- yea there's even now editing of embeds in A111 - i find embeds sometimes MESS shit up too sadly, and i've used amazing embeds XD
How on earth do you edit embeds?
There's a new ting you can use in webUI to mash or merge i dont know how it works XD
https://github.com/klimaleksus/stable-diffusion-webui-embedding-merge
THIS.
THIS IS WHAT WE MEANT LMAO
We figured it out while exhausted at like 12am
AND NOW we'll never go back to training them XD
We didn't MAKE the thing to make them just we knew they existed but forgot to get them XD