HAMP-LIC: Hessian-Aware Mixed-Precision Post-Training Quantization for Learned Image Compression
HAMP-LIC is a Hessian-aware mixed-precision post-training quantization framework for learned image compression models. It uses a four-stage optimization strategy: block-wise sensitivity estimation via Hessian trace, task-aware refinement considering quantization distortion and rate-distortion performance, bit-width allocation under a global model-size constraint, and block-wise reconstruction with a small calibration set.
Learned image compression models achieve strong rate-distortion performance but face high computational complexity and encoding-decoding mismatches across heterogeneous hardware. Uniform fixed-precision quantization reduces these issues but causes severe quality degradation at low bit widths because it ignores per-layer quantization sensitivity differences. HAMP-LIC addresses this by estimating block-wise sensitivity from the Hessian trace, refining sensitivities with a task-aware module, allocating bit widths under a global model-size constraint, and using block-wise reconstruction to suppress quantization error.
The method leverages second-order Hessian trace information to capture layer-wise quantization sensitivity, then refines it with a task-aware module that jointly considers quantization distortion and rate-distortion performance. This enables mixed-precision bit allocation under a global model-size constraint, followed by block-wise reconstruction using a small calibration set to further reduce quantization error.
Efficient low-bit deployment of learned image compression models can reduce computational cost and improve cross-platform consistency, which is relevant for edge devices and heterogeneous hardware environments. The approach may lower barriers to deploying high-performance learned compression in production systems.
The framework could enable more efficient deployment of learned image compression models on resource-constrained hardware, potentially reducing inference costs and improving compatibility across devices.
Observable next signals include publication of full experimental results, comparisons against uniform quantization baselines, and potential adoption or replication by industry groups working on learned image compression for edge deployment.