This contribution reports the EE1-4.1.3 and EE1-4.2.3 results. In both experiments, JPEG-AI is used as an external coder, which feeds I-frames to VTM using the hybrid multilayer framework is available in NNVC16. In EE1-4.1.3, residual coding on top of the prediction for end-to-end AI coded reference picture is enabled. In EE1-4.2.3, skip mode is forced (by setting the QP of Intra frames to the maximum, thus the reconstruction is identical to the externally end-to-end AI coded frame. The complexity of JPEG-AI decoder is 22 kMACs per pixel and has about 5M parameters. BD-rate performance and complexity vs. NNVC15 (with all NN tools off, identical to VTM 23.13) are the following: