Why inpaint works so well. The non-mask area is perfect for image reconstruction

[hl123-123]

Especially since the text reconstruction in the top left corner of the image is so good, it's hard for me to understand. I first thought that the non-mask area of the generated image was directly copied from the non-mask area of the input image. But I can't find the code to merge the original image of the non-mask area with the generated image mask area.

I was very confused. I used autoencoderkl for image reconstruction, and after encode and decode the images directly, I could not carry out extremely detailed image reconstruction. Even if the features in the middle of the non-mask area are constantly controlled by img_condition, such a good image reconstruction effect cannot be achieved