Coding efficiency may be improved by subdividing a block into smaller sub-blocks for prediction. A first rate distortion value of a block optionally partitioned into smaller prediction sub-blocks of a first size is calculated using respective inter prediction modes and transforms of the first size. The residuals are used to encode the block using a transform of a second size smaller than the first size, generating a second rate distortion value. The values are compared to determine whether coding efficiency gains may result from inter predicting the smaller, second size sub-blocks. If so, the block is encoded by generating prediction residuals for the second size sub-blocks, and neighboring sub-blocks are grouped, where possible, based on common motion information. Each resulting composite residual block is transformed by a transform of the same size to generate another rate distortion value. The encoded block with the lowest rate distortion value is used.