Experimenting with training on Anima. I don't think Civitai trainer works very well for this model but YMMV.
Post some gens under this, would ya?
Edit: After some more testing, this seems to be most "accurate" under base Anima (there might be better tuned base models for pixel art, who is to say)
The green background seems to be baked in even though I did tag "green background" in the training data - but I suppose if you're using this for practical purposes you want to manually edit your output anyway.
Description
Comments (9)
I'm assuming you used sprite rips for this and didn't tag your dataset with "green background", so now it's heavily baked in and cannot be removed.
I did tag it with green background, I'm just not very through with prompt testing
Well, tested it with other background colors prompt and yeah the green is super baked in, but my dataset DID tag green background properly so I dunno.
@Selenusuka There are some training methods that can use an alpha mask from a single color background, like AlphaChanelAsMask.
@SugarEnjoyer I only have access to Civit's trainer, can I use that there?
@Selenusuka no, unfortunately you need access to a local training program like OneTrainer or KohyaSS and a decent GPU.
When training models its important to also put in negative or counter examples for as much as possible so it understands what the words mean. So you did indicate green background, but you need to add a bunch of images with black, white, blue, and so on color backgrounds so it understands what you meant with green backgrond and prevents baking in.
for future tries I'll vary the background color of my source so this doesn't happen hopefully
@Selenusuka this should easily fix it. itll be okay