Exploring Latent Cross-Channel Embedding for Accurate 3d Human Pose Reconstruction in a Diffusion Framework.