HAMP-LIC: Hessian-Aware Mixed-Precision PTQ for Learned Image Compression
A new framework named HAMP-LIC has been developed by researchers for mixed-precision post-training quantization (PTQ) of learned image compression (LIC) models. This innovative method tackles the challenges posed by computational complexity and hardware diversity that impede LIC implementation. By utilizing Hessian trace to estimate block-wise sensitivity, it assesses the second-order significance of layers. Additionally, a task-aware refinement module fine-tunes these sensitivities by taking into account both quantization distortion and rate-distortion performance. The four-stage optimization process facilitates the effective and precise low-bit deployment of pretrained LIC models. This research, accessible on arXiv (2608.12239), seeks to reduce quality loss at lower bit widths by considering layer-specific quantization sensitivities, addressing a drawback of uniform fixed-precision quantization.
Key facts
- HAMP-LIC is a Hessian-aware mixed-precision post-training quantization framework for learned image compression.
- It uses a four-stage optimization strategy.
- Block-wise sensitivity is estimated from the Hessian trace to capture second-order importance.
- A task-aware refinement module adjusts sensitivities by jointly considering quantization distortion and rate-distortion performance.
- The method aims to enable efficient and accurate low-bit deployment of pretrained LIC models.
- Uniform fixed-precision quantization suffers severe quality degradation at low bit widths.
- The work is available on arXiv with identifier 2608.12239.
- The paper addresses computational complexity and encoding-decoding mismatches across heterogeneous hardware platforms.
Entities
Institutions
- arXiv