how to train a inpaint controlnet model
I tested your ctrl model before, and found it quite interesting, but there seems to be a little problem in the effect.I want to try to train a model of inpaint control by myself. You can teach me how to train.
Can you give me some manuals or tell me how to do it?
Chengzhiyuan is right I used the script provided by the controlnet library:
Link
I tried different ways of tuning the script bu the original one is still the one which gave the best results.
I build a dataset from lexica images which are all AI generated and might be the explanation for not having enough real photography (dataset was also limited, between a few thousands and ten thousands images).
Just to add another clarification, it is a simple controlnet, this is why the image to inpaint is provided as the controlnet input and not just a mask, I have no idea how to train an inpaint controlnet which would work by just giving a mask to the controlnet and work on an img2img pipeline
I tried,but tested not well
https://huggingface.co/datasets/starinskycc/mytest/tree/main
The experiment was conducted using the dataset above