METHOD OF SYNTHESIZING VIDEO FROM INPUT FRAME USING AUTOREGRESSIVE METHOD, USER ELECTRONIC DEVICE AND COMPUTER-READABLE MEDIUM FOR REALIZING SAID METHOD Russian patent published in 2024 - IPC G06V10/10 G06T11/00 G06T15/00 

FIELD: processing of video data.

SUBSTANCE: invention relates to autoregression synthesis of high-quality video from one frame (image) using a fully convolutional diffusion model. Method of video synthesis by autoregressive method comprises: prediction using trained diffusion model of transition from frame to next frame of next frame based on input frame with simultaneous conditioning of motion between frames of synthesized video using common hidden code m of motion, selected in accordance with a sampling scheme of a hidden motion code based on k-nearest neighbors (kNN). Proposed technical solutions are also suitable for operation on devices with limited processor/memory resources, at least due to the fact that layers of self-attention, initially used in diffusion-probabilistic models of noise elimination (DDPM), are replaced by conditioning by common hidden code of movement.

EFFECT: obtaining video of higher resolution.

14 cl, 4 dwg, 2 tbl