Instance-Adaptive Neural Compression for Low-Bitrate Video Quality

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional image and video coding techniques result in artifacts during decoding and fail to efficiently adapt to varying content, resolutions, and frame rates, placing a significant burden on communication networks and devices due to high data volumes.

Innovation Solution

Utilizing machine learning systems, particularly neural networks, to perform instance-adaptive compression by fine-tuning models for specific data, reducing the need to transmit full model parameters and minimizing bitrate through rate-distortion autoencoders (RD-AEs) with model priors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If conventional video coding techniques are used to compress video data, then the bitrate is reduced, but video quality deteriorates with artifacts appearing during decoding

Engineering Contradiction:
ImprovebitrateVSAvoidvideo quality
Core Design Contradiction:
Quantity of substanceVSManufacturing precision

Solution Approach 1:

The patent changes the parameters of the compression model by fine-tuning neural network weights and parameters specifically for the input video content. This instance-adaptive approach adjusts the compression parameters based on the actual video characteristics, allowing optimal compression ratios while maintaining video quality without the artifacts produced by conventional fixed-parameter coding techniques.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The system performs preliminary fine-tuning of the compression model on the encoder side before actual compression. By pre-adapting the model parameters to the specific video content in advance, the system establishes optimized compression parameters that will maintain video quality during the subsequent compression process, preventing quality deterioration rather than correcting it afterward.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If conventional video coding techniques are used, then compression is achieved, but the system fails to adapt to varying content, resolutions, and frame rates efficiently

Engineering Contradiction:
Improveadaptation to varying contentVSAvoidcompression efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent implements a dynamic compression system where the neural network model parameters are fine-tuned adaptively based on the actual video content characteristics, resolution, and frame rate. This dynamic adaptation allows the system to optimize compression efficiency for each specific video instance rather than using fixed parameters, achieving both high adaptability and maintained productivity through automated instance-adaptive fine-tuning.

Inventive Principle:
Principle #15Dynamics

3Manufacturing precision

If full model parameters are transmitted to achieve high-quality compression, then video quality is maintained, but network bandwidth consumption increases significantly

Engineering Contradiction:
Improvecompression qualityVSAvoiddata volume
Core Design Contradiction:
Manufacturing precisionVSQuantity of substance

Solution Approach 1:

The patent extracts and transmits only the essential fine-tuned parameters and compression data from the full model, rather than transmitting complete model parameters. By selecting and transmitting only the necessary updated parameters that contribute to high-quality compression, the system maintains video quality while significantly reducing the data volume transmitted over the network.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system performs preliminary fine-tuning on the encoder side and prepares the compressed data with embedded necessary parameters before transmission. By pre-processing and packaging only the essential information needed for reconstruction, the system avoids transmitting redundant full model parameters, thus maintaining compression quality while minimizing network bandwidth consumption.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4218238B1Instance-adaptive image and video compression using machine learning systems
Publication Date: 2025.12.24 QUALCOMM INC
  • EP4218238B1 patent drawingFigure 1
  • EP4218238B1 patent drawingFigure 2A~2C
  • EP4218238B1 patent drawingFigure 2D

AI summary

Techniques are described for compressing data using machine learning systems and tuning machine learning systems for compressing the data. An example process can include receiving, by a neural network compression system (e.g., trained on a training dataset), input data for compression by the neural network compression system. The process can include determining a set of updates for the neural network compression system, the set of updates including updated model parameters tuned using the input data. The process can include generating, by the neural network compression system using a latent prior, a first bitstream including a compressed version of the input data. The process can further include generating, by the neural network compression system using the latent prior and a model prior, a second bitstream including a compressed version of the updated model parameters. The process can include outputting the first bitstream and the second bitstream for transmission to a receiver.