Reconstructed prediction errors in a sub-portion (12) of a residual block (11) obtained by inverse transforming a transform block (10) are modified in order to spatially improve localized portions of the residual block (11), e.g. to compensate for visual artifacts from transform coding. The modification affects reconstructed prediction errors in the sub-portion (12) of the residual block (11) but not reconstructed prediction errors in a remaining portion of the residual block (11).