Sudoku, the popular logic puzzle, would have a greater artistic appeal if the final completed puzzle could be transformed into an image similar to nonogram outputs. To do this, we solve a mixed-integer nonlinear programming problem (MINLP) to find the Sudoku configuration that most closely represents a target image via an integer-colour mapping. This alone produces inadequate results, so we relax the problem by adding transparent regions to the target image and removing certain MINLP constraints. To produce puzzles that are feasible and interesting to players, additional components are added to avoid tedious recounting and unnecessary colouring.