Techniques are described for compressing and decompressing data using machine learning systems. An example process can include receiving a plurality of images for compression by a neural network compression system. The process can include determining, based on a first image from the plurality of images, a first plurality of weight values associated with a first model of the neural network compression system. The process can include generating a first bitstream comprising a compressed version of the first plurality of weight values. The process can include outputting the first bitstream for transmission to a receiver.

Motion compensation along different directions is disclosed. A method for video processing including: determining, for a conversion between a current video block of a video and a bitstream representation of the current video block, optical flow associated with the current video block in an optical flow-based motion refinement process or prediction process, wherein the optical flow is derived along directions that are different from a horizontal direction and/or a vertical direction; and performing the conversion based on the optical flow.

A method for video processing includes: refining motion information of a video block by using a multi-step refinement processing, multiple refined motion vectors (MVs) of the video block being derived iteratively in respective steps of the multi-step refinement processing, and performing a video processing on the video block based on the multiple refined MVs of the video block.

An image encoding/decoding method and apparatus are provided. An image decoding method according to the present disclosure is performed by an image decoding apparatus. The image decoding method comprises constructing a merge candidate list for a current block based on a prediction mode of the current block, deriving motion information of the current block based on the merge candidate list, and generating a prediction block of the current block based on the motion information. Information on a maximum number of merge candidates included in the merge candidate list may be obtained through a sequence parameter set, and, based on the prediction mode being a subblock-based merge mode, the maximum number of merge candidates may be determined based on whether an affine mode is available for the current block.

An image encoding/decoding method and apparatus are provided. An image decoding method according to the present disclosure is performed by an image decoding apparatus. The image decoding method may comprise deriving a prediction sample of a current block based on motion information of the current block, determining whether prediction refinement with optical flow (PROF) applies to the current block, deriving, based on that the PROF applies to the current block, a difference motion vector for each sample position in the current block, deriving a gradient for each sample position in the current block, deriving a PROF offset based on the difference motion vector and the gradient, and deriving a refined prediction sample for the current block based on the PROF offset.

An example method includes determining, without receiving explicit signaling, whether motion compensation for a current block of a current picture of video data is to be performed using a four-parameter affine motion model (AMM) defined by two motion vectors (MVs) or using a six-parameter AMM defined by three MVs; deriving values of predictors for MVs of the AMM of the current block; decoding a representation of differences between the values of the MVs of the AMM for the current block and the values of the predictors; determining the values of the MVs of the AMM for the current block from the values of the predictors and the decoded differences; determining, based on the determined values of the MVs of the AMM for the current block of video data, a predictor block of video data; and reconstructing the current block based on the predictor block.

An example method includes determining, without receiving explicit signaling, whether motion compensation for a current block of a current picture of video data is to be performed using a four-parameter affine motion model (AMM) defined by two motion vectors (MVs) or using a six-parameter AMM defined by three MVs; deriving values of predictors for MVs of the AMM of the current block; decoding a representation of differences between the values of the MVs of the AMM for the current block and the values of the predictors; determining the values of the MVs of the AMM for the current block from the values of the predictors and the decoded differences; determining, based on the determined values of the MVs of the AMM for the current block of video data, a predictor block of video data; and reconstructing the current block based on the predictor block.

A method of motion compensation for geometry representation of 3D data is described herein. The method performs motion compensation by first identifying correspondent 3D surfaces in time domain, then followed by a 3D to 2D projection of motion compensated 3D surface patches, and then finally performing 2D motion compensation on the projected 3D surface patches.

A method of motion compensation for geometry representation of 3D data is described herein. The method performs motion compensation by first identifying correspondent 3D surfaces in time domain, then followed by a 3D to 2D projection of motion compensated 3D surface patches, and then finally performing 2D motion compensation on the projected 3D surface patches.

Encoding or decoding picture information can involve determining a coding mode associated with a current coding unit including picture information; determining, based on the coding mode, a first precision level associated with a first portion of a motion vector information associated with the current coding unit, and a second precision level associated with a second portion of the motion vector information; obtaining a motion vector associated with the current coding unit based on the motion vector information and the first and second precision levels; and encoding or decoding at least a portion of the picture information included in the current coding unit based on the coding mode and the motion vector.