Granted Patent B2

US 12,739,412 · App. 18/421,809 · Granted Sep 15, 2026

Inventors: Ziyad A. Ibrahim (Redmond, WA); Sudhakar V. Prabhu (Redmond, WA)

Assignee: Microsoft Technology Licensing, LLC

H04N19/33H04N19/31H04N19/42H04N19/503H04N19/593H04N19/70

Abstract

Claims (82)

1 . A computer system comprising a processor system and memory, wherein the computer system is configured to perform operations to compress media using a generative artificial intelligence (“AI”) model, the operations comprising:

receiving a first version of input media, the first version having a first resolution;

converting the first version to a second version of the input media, the second version having a second resolution lower than the first resolution;

compressing the second version according to a media compression format, thereby producing compressed data for the second version;

outputting the compressed data for the second version;

receiving input tokens that represent input syntax elements, respectively, of the compressed data for the second version, wherein the input tokens are encoded in an input format for the generative AI model;

providing, to the generative AI model, the input tokens;

receiving, from the generative AI model, predicted tokens that represent output syntax elements, respectively, of compressed data for output media at the first resolution;

reconstructing the output media from the predicted tokens;

determining differences between the first version and the output media;

compressing the differences, thereby producing compressed data for the differences; and

outputting the compressed data for the differences for decompression as an augmentation stream.

2 . The computer system of claim 1 , wherein the second resolution is lower than the first resolution in terms of quality, bitrate, spatial resolution, and/or temporal resolution.

3 . The computer system of claim 1 , wherein the operations further comprise, with the generative AI model, processing the input tokens to determine the predicted tokens, including:

converting the input tokens into encoder input embedding vectors;

determining, based on the encoder input embedding vectors, encoder output embedding vectors using multiple layers of an encoder of the generative AI model;

determining, based on decoder input embedding vectors and the encoder output embedding vectors, decoder output embedding vectors using multiple layers of a decoder of the generative AI model, wherein the decoder input embedding vectors are based on prior predicted tokens; and

converting the decoder output embedding vectors into the predicted tokens.

4 . The computer system of claim 1 , wherein the determining the differences includes:

determining pixel value differences between pixel values of the first version and corresponding pixel values of the output media, wherein the corresponding pixel values are spatial collocated pixel values in temporally corresponding frames.

5 . The computer system of claim 1 , wherein the operations further comprise:

multiplexing, according to a container format, the compressed data for the second version and the compressed data for the differences, wherein the compressed data for the second version and the compressed data for the differences are output as part of containers in the container format.

6 . In a computer system that implements a generative artificial intelligence (“AI”) model, a method of training the generative AI model to compress media, the method comprising:

identifying a set of training data; and

training the generative AI model in multiple training iterations using the set of training data, including, in a given training iteration of the multiple training iterations:

receiving first input tokens that represent first input syntax elements, respectively, of compressed data for a first version of input media, the first version having a first resolution, the first version having been compressed according to a media compression format to produce the compressed data for the first version, wherein the first input tokens are encoded in an input format for the generative AI model;

receiving second input tokens that represent second input syntax elements, respectively, of compressed data for a second version of the input media, the second version having a second resolution lower than the first resolution, the second version having been compressed according to the media compression format to produce the compressed data for the second version of the input media, wherein the second input tokens are encoded in the input format for the generative AI model;

providing, to the generative AI model, the first input tokens and the second input tokens;

receiving, from the generative AI model, predicted tokens that represent output syntax elements, respectively, of compressed data for output media at the first resolution;

determining a measure of loss based at least in part on the predicted tokens; and

updating one or more parameters of the generative AI model based at least in part on the measure of loss.

7 . The method of claim 6 , wherein the set of training data includes, for each of multiple examples of input media, first input tokens that represent first input syntax elements, respectively, of compressed data for a first version, having the first resolution, of that example of input media and second input tokens that represent second input syntax elements, respectively, of compressed data for a second version, having the second resolution, of that example of input media, and wherein, for each of the multiple examples of input media, the first version and the second version of that example of input media have been compressed according to the media compression format, and wherein, for each of the multiple examples of input media:

the first version and the second version of that example have been compressed using a common set of compression settings and a common profile of the media compression format;

the first version and the second version of that example differ, in terms of quality, bitrate, spatial resolution, and/or temporal resolution, by a common scaling factor for that attribute; and

is part of a given genre of media.

8 . The method of claim 6 , further comprising:

receiving the compressed data for the first version;

partially decompressing the compressed data for the first version, thereby producing the first input syntax elements according to the media compression format;

converting the first input syntax elements into the first input tokens;

receiving the compressed data for the second version;

partially decompressing the compressed data for the second version, thereby producing the second input syntax elements according to the media compression format; and

converting the second input syntax elements into the second input tokens.

9 . The method of claim 6 , further comprising, with the generative AI model, processing the first input tokens and the second input tokens to determine the predicted tokens, including:

converting the second input tokens into encoder input embedding vectors;

determining, based on the encoder input embedding vectors, encoder output embedding vectors using multiple layers of an encoder of the generative AI model;

converting the first input tokens into decoder input embedding vectors;

determining, based on the decoder input embedding vectors and the encoder output embedding vectors, decoder output embedding vectors using multiple layers of a decoder of the generative AI model; and

converting the decoder output embedding vectors into the predicted tokens.

10 . The method of claim 6 , wherein the determining the measure of loss includes:

determining a measure of conformity of the predicted tokens to syntax of the media compression format, wherein the measure of conformity of the predicted tokens to syntax of the media compression format quantifies loss in terms of deviations from the syntax of the media compression format.

11 . The method of claim 6 , wherein the training further includes, in the given training iteration, reconstructing the output media from the predicted tokens, and wherein the determining the measure of loss includes:

determining a measure of quality degradation of the output media relative to the first version of the input media.

12 . The method of claim 11 , wherein the measure of quality degradation is peak signal-to-noise ratio, structural similarity index, multi-scale structural similarity index, or video multi-method assessment fusion.

13 . The method of claim 6 , wherein the parameters of the generative AI model include one or more of:

an embedding matrix for determining decoder input embedding vectors from the first input tokens, determining encoder input embedding vectors from the second input tokens, and/or for determining the predicted tokens from output embedding vectors;

linear projections for determining inputs to multi-head attention sub-layers of the generative AI model;

weights and offsets for feed-forward neural networks of the generative AI model; and

parameters of a softmax function of the generative AI model.

14 . One or more non-transitory computer-readable media having stored thereon computer-executable instructions for causing a processor system, when programmed thereby, to perform operations to decompress media using a generative artificial intelligence (“AI”) model, the operations comprising:

receiving input tokens that represent input syntax elements, respectively, of compressed data for a second version of input media, the second version having been compressed according to a media compression format to produce the compressed data for the second version, wherein the second version approximates a first version of the input media, the first version having a first resolution, the second version having a second resolution lower than the first resolution, and wherein the input tokens are encoded in an input format for the generative AI model;

providing, to the generative AI model, the input tokens;

receiving, from the generative AI model, predicted tokens that represent output syntax elements, respectively, of compressed data for output media at the first resolution; and

reconstructing the output media from the predicted tokens.

15 . The one or more non-transitory computer-readable media of claim 14 , wherein the second resolution is lower than the first resolution in terms of quality, bitrate, spatial resolution, and/or temporal resolution.

16 . The one or more non-transitory computer-readable media of claim 14 , wherein the operations further comprise:

receiving the compressed data for the second version;

partially decompressing the compressed data for the second version, thereby producing the input syntax elements according to the media compression format; and

converting the input syntax elements into the input tokens.

17 . The one or more non-transitory computer-readable media of claim 14 , wherein the operations further comprise, with the generative AI model, processing the input tokens to determine the predicted tokens, including:

converting the input tokens into encoder input embedding vectors;

determining, based on the encoder input embedding vectors, encoder output embedding vectors using multiple layers of an encoder of the generative AI model;

determining, based on decoder input embedding vectors and the encoder output embedding vectors, decoder output embedding vectors using multiple layers of a decoder of the generative AI model, wherein the decoder input embedding vectors are based on prior predicted tokens; and

converting the decoder output embedding vectors into the predicted tokens.

18 . The one or more non-transitory computer-readable media of claim 14 , wherein the reconstructing the output media from the predicted tokens includes:

converting the predicted tokens to the output syntax elements, respectively, in the media compression format; and

decompressing the output syntax elements using a media decoder for the media compression format.

19 . The one or more non-transitory computer-readable media of claim 14 , wherein the operations further comprise:

receiving compressed data for differences between the first version and the output media;

decompressing the compressed data for the differences, thereby reconstructing a version of the differences; and

combining the reconstructed version of the differences with the output media.

20 . The one or more non-transitory computer-readable media of claim 19 , wherein the operations further comprise:

demultiplexing, according to a container format, the compressed data for the second version and the compressed data for the differences, wherein the compressed data for the second version and the compressed data for the differences are received as part of containers in the container format.

Assignments (1)

Continuity (1)

References Cited (23)