Sky replacement

A method comprising:

• generating a first region mask, a second region mask, and a third region mask using a same machine learning model and corresponding to a same semantically related region of a first image;
• generating a defringing layer by combining the first region mask with grayscale version of a second image;
• generating a region-specific layer by combining the second region mask and the third region mask to produce a combined region mask, and combining the combined region mask with the second image;
• and
• generating a composite image by combining a plurality of layers including the first image, the defringing layer, and the region-specific layer, wherein the composite image is generated based on a first user input to adjust the plurality of layers locally and a second user input to adjust the plurality of layers globally.

An apparatus comprising:

• a mask generation network configured to generate a plurality of region masks based on a first image using a same mask generation network, wherein the plurality of region masks include a first region mask and a second region mask, and wherein the first region mask corresponds to a same semantically related region of the first image as the second region mask and has a more gradual mask boundary than the second region mask;
• a defringing component configured to generate a defringing layer by combining the first region mask of the plurality of region masks with grayscale version of a second image;
• a region-specific layer component configured to generate a region-specific layer by combining the second region mask of the plurality of region masks with the second image;
• and
• a layer composition component configured to generate a composite image by combining a plurality of layers including the first image, the defringing layer, and the region-specific layer, wherein the composite image is generated based on a first user input to adjust the plurality of layers locally and a second user input to adjust the plurality of layers globally.