In this contribution it is proposed to modify the size of the filter kernel that is used to estimate HoG in DIMD based on the size of a block being predicted. For smaller blocks, a smaller kernel size is used. On top of ECM-13.0, the following BD-rate gain results and encoding/decoding runtimes are reportedly obtained {Y/U/V/EncT/DecT}: {-0.%, -0.%, -0.%, %, 100%} for AI configuration. The proposed method does not increase the encoder or decoder runtime.