Exemplary video processing methods and apparatuses for encoding or decoding a current block by inter prediction are disclosed. Input data of a current block is received and partitioned into sub-partitions and motion refinement is independently performed on each sub-partition. A reference block for each sub-partition is obtained from one or more reference pictures according to an initial motion vector (MV). A refined MV for each sub-partition is derived by searching around the initial MV with N-pixel refinement. One or more boundary pixels of the reference block for a sub-partition is padded for motion compensation of the sub-partition. A final predictor for the current block is generated by performing motion compensation for each sub-partition according to its refined MV. The current block is then encoded or decoded according to the final predictor.