The invention refers to a method for reducing noise for coding of noisy images or image sequences. The invention comprises the steps of adding a prediction error (e q [x, y]) and a predicted frame (f [x, y]) or a predicted block for receiving a decoded frame (g q [x, y]) or a decoded block to be further used in a prediction loop by an encoder (ENC) or to be sent to the output of a decoder (DEC), wherein the reference frame (g q [x, y]) or the reference block (FR) consist of a respective useful signal part and a noise signal part. Furthermore, it comprises the step of denoising of the reference frame (g q [x, y]) or reference block (FR) only with a dedicated noise reducing filter (WF) to reduce or eliminate the noise signal part of the reference frame (g q [x, y]) or reference block (FR).