Photorealistic detail manipulation via multi-light images

In this paper, a new image synthesis framework is proposed to provide delicate description of a scene. The input is a set of images that capture the same scene under different lighting conditions. By solving a gradient domain optimization problem, the details of the input images can be combined together into an auxiliary image. Then a quadratic filter is used to divide the auxiliary image into base layer and detail layer. To make the contents in the detail layer more visible, the local contrast of the detail layer is enhanced. The base layer is replaced by one of the input image with a spatial varying exposure function to adjust the extreme luminance in base layer. The final desired image is generated by multiplying the enhanced detail layer to the adjusted base layer. It is shown that the final synthesized image is able to provide a more detailed description of the captured scene than anyone of the input images.