Do you expand all convolutional layers of UNet to 13 channels initialized with zero weights or only expand the first convolutional layer of UNet to 13 channels? Do you use the pre-trained SDXL inpainting models to initialize the denoiser inpainting Unet?
Thanks for your excellent work.
Do you expand all convolutional layers of UNet to 13 channels initialized with zero weights or only expand the first convolutional layer of UNet to 13 channels? Do you use the pre-trained SDXL inpainting models to initialize the denoiser inpainting Unet?