Although radiofrequency ablation has become a worldwide accepted curative therapy for the treatment of cardiac arrhythmias, there is still a lack of lesion monitoring during ablation procedure. Magnetic Resonance (MR) thermometry based on the water proton frequency could give relevant information on local tissue temperature increase but requires dealing with motion and susceptibility artifacts due to the cardiac and respiratory motion. A real-time framework that overcomes these difficulties is described in this paper. The proposed approaches address both inter-scan and inplane motion using a rapid MR-acquisition sequence coupled with a robust optical-flow registration of all incoming images. The strategy was found to be robust and suitable to follow the temperature evolution in 10 healthy volunteers under free breathing conditions with a temperature standard deviation of 2.2 °C in the cardiac muscle at an update rate of approximately 1 Hz. Since local temperature rises due to RF ablation can lead to T1 and T2 relaxation time local variations, and thus to errors in the image-based registration process, we also analyze the performance of the algorithm during radiofrequency ablation applied under well controlled experimental conditions (no respiratory motion and cardiac triggered) using an ex vivo perfused pig heart. Contact electrophysiological signals were recorded simultaneously with magnetic resonance imaging by combining hardware and software filtering. A working heart setup, specifically designed for this project, provided a relevant model to assess the quality of cardiac MR-thermometry.