DiRIC: Diffusion Prior Refinement for Efficient Low-rate Image Compression

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Low bitrate image reconstruction is challenging.
• Diffusion models offer prior knowledge but struggle with semantic integration in compression.
• Existing methods lack alignment between diffusion priors and compressor objectives.

Purpose of the Study:

• To propose Diffusion Prior Refinement for Efficient Low-rate Image Compression (DiRIC).
• To leverage Stable Diffusion for high-quality, efficient low-rate image compression.
• To improve semantic feature integration and decoding efficiency.

Main Methods:

• Utilizing a multi-feature compressor for low-level information at low bitrates.
• Employing a pre-embedding mechanism for robust hybrid semantics.
• Designing a consistency skip module to refine diffusion priors.
• Implementing a noise-level estimator to reduce decoding sampling steps.

Main Results:

• Achieved state-of-the-art perceptual fidelity.
• Significantly outperformed previous methods in statistical fidelity (FID and PSNR).
• Demonstrated a 147.44% BD Rate improvement in FID and 84.63% in PSNR compared to SoTA diffusion baselines.
• Achieved a 19x increase in decoding speed.

Conclusions:

• DiRIC effectively reconstructs high-quality images at low bitrates.
• The method integrates semantic information and diffusion priors efficiently.
• DiRIC offers superior perceptual and statistical fidelity with enhanced decoding speed.