For very large images, you'll need to construct an image pyramid, which is a hierarchical set of images, each of which is an increasingly lower resolution version of the original image. Each image in the pyramid is subdivided into tiles, so that only the portions in view need to be loaded. Google Earth calculates the current viewpoint and loads the tiles that are appropriate to the user's distance from the image. As the viewpoint moves closer to the PhotoOverlay, Google Earth loads higher resolution tiles. Since all the pixels in the original image can't be viewed on the screen at once, this preprocessing allows Google Earth to achieve maximum performance because it loads only the portions of the image that are in view, and only the pixel details that can be discerned by the user at the current viewpoint.
When you specify an image pyramid, you also modify the in the element to include specifications for which tiles to load.