SIFT flow is a method to align an image to its neighbors in a large image collection consisting of a variety of scenes. Originally proposed for cross-scene alignment, it has good potential for robust dense matching between images of a scene taken under large viewing condition changes. However, this potential has not been fully exploited due to dense SIFT in SIFT flow is extracted under single pre-defined scale. In this paper, we explore this potential and propose new algorithms to select proper scales while applying SIFT flow for dense image matching. By studying the behavior of SIFT flow under scale changes, we propose two new concepts, namely, the optimal relative scale factor (ORSF) and the optimal matching scale (OMS) using gradient-enhanced normalized mutual information. ORSF and OMS define the most proper scales during SIFT flow matching. Suboptimal and optimal methods for estimating ORSF and OMS are proposed. It is shown that by applying ORSF and OMS, the accuracy of SIFT flow for dense image matching is greatly improved on images with significant scale changes.