some questions about the paper

[dadwadw233]

[dadwadw233]

i have some questions about in/out-painting inplementation details from the paper. you mentioned that the in/out-painting model need Cs+1 extra input channels and you write down the operation method “Concat(m,mSt,St)” , i think St and mSt have some same info? why can‘t we just concat the noised data and in/out-painting mask？🤔

[djFatNerd]

Hi, sorry for the late reply. I do agree these two are conceptually the same but I didn't test it with out masked data. We followed the implementation from diffuser's training of inpainting models.