Saliency-based collage generation techniques are described. A collage generation module is executed by a computing device and receives multiple digital images. The collage generation module then generates multiple saliency maps from the digital images that describe a visual saliency of respective pixels in the digital images. The saliency maps are then used by the collage generation module to fit bounding boxes to portions of the digital images that are considered salient. Collage candidates are generated by the collage generation module based on predefined layouts and the bounding boxes that are fit the portions of the digital images. The collage generation module then selects at least one collage from these collage candidates for output to a user based on the determined amounts of deviation.