A Noise Constrained Diffusion (NC-Diffusion) Framework for High Fidelity Image Compression
A new AI framework repurposes compression noise to train diffusion models, achieving state-of-the-art results.
A research team including Zhenyu Du, Yanbo Gao, and Shuai Li has published a novel AI framework called Noise Constrained Diffusion (NC-Diffusion) that significantly advances diffusion-based image compression. The core innovation addresses a major flaw in previous approaches: the mismatch between the random Gaussian noise used to train standard diffusion models and the actual quantization noise introduced during image compression. NC-Diffusion cleverly formulates this quantization noise as the starting point for its diffusion process, creating a direct, constrained path from the compressed image back to the original. This alignment drastically improves the model's inference efficiency and reconstruction accuracy.
Beyond the core noise constraint, the framework incorporates two key enhancements for superior image quality. An adaptive frequency-domain filtering module is integrated into the U-Net architecture to strengthen skip connections, specifically enhancing the recovery of crucial high-frequency details often lost in compression. Furthermore, a zero-shot sample-guided enhancement method is employed to further boost the final image fidelity without requiring additional task-specific training. Experiments across multiple benchmark datasets confirm that NC-Diffusion achieves the best performance compared to existing methods, setting a new standard for high-fidelity, AI-powered image compression. The paper has been accepted for publication in the IEEE Transactions on Circuits and Systems for Video Technology.
- Repurposes quantization noise from compression as the diffusion training signal, solving a critical noise mismatch problem.
- Includes an adaptive frequency filter in the U-Net to enhance high-frequency detail recovery in compressed images.
- Demonstrates state-of-the-art performance on benchmarks, outperforming existing diffusion-based and traditional compression methods.
Why It Matters
This enables much higher quality image compression for storage and streaming, potentially reducing bandwidth use without sacrificing visual detail.