This contribution reports the results of EE1-1.2, which uses DCT transformed inputs on top of EE1-1.1. The method applies the DCT to the inputs of the model and the inverse DCT to the outputs. The spatial resolution inside the model is reduced by a factor of four, and the reduced complexity is balanced by increasing the number of channels. The proposed model has a complexity of 16.9 kMAC/pixel and 0.2M parameters, while the model from EE1-1.1 has 17.0 kMAC/pixel and 0.05M parameters. Stage 3 training of the proposed model is performed. The performance of the int16 model (including NNIntra) compared to the NNVC-8.0 anchor (NNIntra+LOP) is reported to be: