ALT Overview of the proposed MoCoDAD. A sequence of N skeletal motions is split into past (top-right X1:k frames) and future (top-left Xk+1:N frames). During training, the Forward Diffusion block adds noise to the future frames, shifting each joint by a random vector displacement of varying intensity (increasing with the diffusion timestep t). Then the Reverse Diffusion learns to estimate the noise. A key aspect of MoCoDAD is the conditioning, i.e. how to encode the past clean k frames and guide the synthesis of relevant futures.

ALT Overview of the proposed MoCoDAD. A sequence of N skeletal motions is split into past (top-right X1:k frames) and future (top-left Xk+1:N frames). During training, the Forward Diffusion block adds noise to the future frames, shifting each joint by a random vector displacement of varying intensity (increasing with the diffusion timestep t). Then the Reverse Diffusion learns to estimate the noise. A key aspect of MoCoDAD is the conditioning, i.e. how to encode the past clean k frames and guide the synthesis of relevant futures.