Inpainting weird colors
I think u may need more real life data instead of synthetic data
it does try to inpaint but the colors are weird
do u have the training code we can use?
As I explain in the other discussion #2 I used the basic control net training script (link also in the other discusson). My dataset was built from lexica images Everett time with both inpainting and outpainting version. For example :
From what I saw, if the prompt describe accurately the image it works quite well but the farther you diverge, the more impact on the global image.
Like putting something in place of the head on a human body 😁.
yeah i agree, but i think for this controlnet it needs an extra channel for the mask so it doesnt mess with the colors of other areas
yeah i agree, but i think for this controlnet it needs an extra channel for the mask so it doesnt mess with the colors of other areas
agreed! would it be possible to implement this?
Current diffusers implementation of controlnet convert the input image to rgb so no it is no possible without developing a new kind of controlnet.