In　this　paper,　we　propose　a　novel　edge　consistency　metric　for　multi-modal　correspondence.　It　is　based　on　a　novel　observation　on　image　truncated　SVD　(singular　value　decomposition)　termed　regression　robustness,　which　describes　the　fact　that,　a　good　approximation　from　image　truncated　SVD　can　be　inherited　even　if　the　eigen-images　change　due　to　expansion　and　channel-dependent　offsets.　Compared　to　state-of-the-arts,　multi-modal　edge　consistency　metric　can　simultaneously　handle　multiple　images　with　complex　modality　changes,　including　local　variation,　gradient　reverse,　intensity　order　change,　and　texture　loss.　Its　complexity　is　almost　linear　to　pixel　number.　Remarkable　accuracies　have　been　achieved　in　experiments.