The approach that I took was to align the images using normalized cross-correlation. I also aligned the images by their edges for one of the bells and whistles. My first implementation scored the alignment of images by normalized the values of each image then taking the dot product of those two images. To ensure the script ran in areasonable time I implemented an image pyramid search procedure. The only image that this method was not able to acceptably align was the emir image. Looking at the NCC-aligned image shows that the blue and green frames are aligned but not the red frame. I belive that this is because the clothes that the man is reflects a large amount of green and blue light but almost no red light. This makes it so that blue and green are more likely to be aligned since they both are concentrated in the same area, whereas there is very little red in the same area.
To find the edges of the images I convolved the images with a Sobel Filter. I then converted the resulting matrix to a boolean matrix: true corresponds to elements with value at least one standard deviation above the average value. I then scored the alignment of the two edge-detected images using the same normalized cross-correlation procedure. For the emir image, this method did significantly better because edge features are less susceptible to different intensities of the different channels because it's features are based on differences between areas rather than raw values.