Diffusion Model Image Compression for Realistic Reconstruction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional image compression methods often produce blurry and unrealistic images, especially at low bitrates, due to their focus on pixel-wise accuracy rather than perceptual quality.

Innovation Solution

A system utilizing a generative model, specifically a diffusion model, on the decoder side to synthesize realistic details while minimizing bit usage, along with a parameter estimation network to optimize the rate-distortion objective.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If pixel-wise accuracy is emphasized to minimize distortion, then measurement precision is improved, but the reconstructed images become blurry and unrealistic

Engineering Contradiction:
Improvepixel-wise accuracyVSAvoidrealism of reconstructed images
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent changes the optimization objective from pixel-wise metrics (MSE) to perceptual quality metrics that evaluate image realism and visual fidelity. This parameter change in the loss function enables the model to prioritize realistic texture synthesis over exact pixel matching, resolving the contradiction between measurement precision and image realism.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces traditional mechanical compression algorithms with a data-driven neural network system that learns optimal compression strategies from training data. This substitution enables the model to balance pixel accuracy and perceptual quality automatically, producing realistic images that traditional methods cannot achieve.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Loss of substance

If compression bitrate is reduced to minimize transmission cost, then loss of substance is improved, but image quality and realism deteriorate

Engineering Contradiction:
Improvetransmission costVSAvoidimage quality
Core Design Contradiction:
Loss of substanceVSReliability

Solution Approach 1:

The patent introduces a neural network model as an intermediary between the compressed bitstream and the final reconstructed image. This intermediary model synthesizes realistic image details from compressed representations, enabling high-quality reconstruction at low bitrates by mediating the transformation from compact data to visually faithful images.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent changes the reconstruction approach from direct pixel prediction to generative modeling that optimizes for perceptual quality. This parameter change in the reconstruction strategy allows the system to maintain image realism even when transmission cost is minimized through aggressive compression.

Inventive Principle:
Principle #35Parameter changes

3Device complexity

If traditional encoder-decoder methods are used to simplify device complexity, then device complexity is reduced, but image realism and perceptual quality are compromised

Engineering Contradiction:
Improvecompression system structureVSAvoidperceptual quality
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The patent replaces traditional handcrafted encoder-decoder architectures with learned neural network models that automatically optimize compression and reconstruction. This substitution improves perceptual quality by enabling the system to adapt to image content and prioritize visually important features, overcoming the limitations of fixed traditional methods.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent creates a universal neural network model that can handle various compression scenarios and image types through learned representations. This multi-functional approach maintains perceptual quality across different conditions without requiring complex specialized systems for each scenario.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4557727A1Lossy image compression with diffusion models
Publication Date: 2025.05.21 DISNEY ENTERPRISES INC
  • EP4557727A1 patent drawingFigure 1
  • EP4557727A1 patent drawingFigure 2
  • EP4557727A1 patent drawingFigure 3

AI summary

In some embodiments, a method receives a quantized latent representation of an image in a latent space. The image is encoded into a representation in the latent space and quantized to generate the quantized latent representation. A time step parameter is received that is generated based on the representation. The method performs an inverse quantization process to generate a reconstructed representation. A diffusion model performs a denoising process for a number of iterations based on the time step parameter to remove noise from the reconstructed representation to generate a denoised reconstructed representation. The denoised reconstructed representation is decoded into a reconstructed image.