Real-time cardiac MRI is of clinical importance in some patients (e.g. with arrythmias or in pediatrics). However, standard real-time MRI suffers from compromised spatiotemporal resolution. A new method for real-time denoising is presented to overcome signal-to-noise ratio (SNR) limitations of highly accelerated TSENSE acquisitions. This method is based on a multi-resolution rigid registration and an extended non-local means filter that uses redundancy between successive frames. The performances of this method, studied in 5 volunteers and 5 patients, and its computational efficiency allow in-line processing of highly accelerated real-time cardiac images, and should be compatible with clinical applications including stress studies.