Spatial models of voting have dominated mathematical political theory since the seminal work of Downs. The Downsian model assumes that each elector votes on the basis of his utility function which depends only on the distance between his preferred policy platform and the ones proposed by candidates. A succession of papers introduces valence issues into the model, i.e., candidates' characteristics which are independent of the platforms they propose. So far, little is known about which of the existing utility functions used in valence models is the most empirically founded. Using a large survey run prior to the 2007 French presidential election, we evaluate and compare several spatial voting models with valence. Existing models perform poorly in fitting the data. However, strong empirical regularities emerge. This leads us to a new model of valence that we call the intensity valence model. This new model makes sense theoretically and is grounded empirically.