This contribution details the architectural improvements on JVET-AC0155 proposed in the EE1-1.3 tests and the corresponding computational complexity and BD-rate changes. Test 1.3.1 reports the baseline performance of the original architecture (from JVET-AC0155) with NNVC-4.0. Test 1.3.2a removes the redundant spatial processing of constant inputs (e.g. QP). Test 1.3.2b removes the partitioning input from the intra-frame models. Test 1.3.3 unifies the intra- and inter-frame models for luma and chroma, respectively. Test 1.3.5a decomposes expensive 3x3 convolutions into separable 3x1 and 1x3 convolutions. Test 1.3.5b proposes a model combining simplifications from the above sub-tests. The training and inference cross-check was performed for tests 1.3.1 and 1.3.5.