DynaQuant: Dynamic Mixed-Precision Quantization for Learned Image Compression

The recent publication of DynaQuant marks a pivotal advancement in learned image compression by addressing the inefficiencies of static, uniform bit-width quantization methods. Traditional techniques often lead to suboptimal performance due to their inability to adapt to diverse data distributions. DynaQuant's innovative approach combines content-aware quantization with a dynamic bit-width selector, allowing for a more tailored response to the statistical variations of latent features. This dual-level adaptation is facilitated by a Distance-aware Gradient Modulator (DGM), which enhances the learning process beyond conventional methods. Experimental results indicate that DynaQuant achieves rate-distortion performance on par with full-precision models while significantly lowering computational and storage requirements. This development not only improves the efficiency of image compression but also opens new avenues for applying learned compression techniques in various fields, making it …