HAMP‑LIC: Optimizing Learned Image Compression with Hessian‑Aware Post‑Training Quantization

Challenges in Learned Image Compression

Deep‑learning based image compression (LIC) models deliver strong rate‑distortion performance, yet they still face two core challenges when deployed on heterogeneous hardware. First, the high computational complexity makes encoding and decoding slow, especially on edge devices. Second, when a model is quantized to a fixed precision, the varying sensitivity of each layer leads to significant reconstruction errors, causing quality discrepancies across platforms. Consequently, many teams avoid deploying LIC in production due to uncertainty in quality and efficiency. Deploying LIC in mobile apps, digital cameras, and video streaming demands solutions that balance visual quality with limited resources.

HAMP‑LIC: Core Concept

HAMP‑LIC (Hessian‑Aware Mixed‑Precision Post‑Training Quantization for Learned Image Compression) tackles both problems simultaneously. It leverages Hessian information—the second derivative of the loss—to assess the sensitivity of each network block. This allows HAMP‑LIC to assign higher precision to blocks that heavily influence reconstruction quality, while less critical blocks can be compressed more aggressively without large quality loss. This contrasts with uniform fixed‑precision quantization that assumes all layers have equal importance.

Four‑Stage Optimization Pipeline

Using the Hessian trace, HAMP‑LIC computes the contribution of each block to loss change, providing a more accurate second‑order importance than first‑order gradient‑based methods.

The initial sensitivities are adjusted by jointly considering quantization distortion and rate‑distortion performance, ensuring that precision allocation aligns with the ultimate compression goal.

Guided by the refined sensitivity profile, the algorithm assigns bit widths to each block while respecting a global model‑size limit, balancing storage efficiency and reconstruction quality.

A tiny calibration set fine‑tunes each block’s output, reducing residual quantization error and eliminating cross‑platform encoding‑decoding mismatches.

Practical Impact on Edge Devices

Applying HAMP‑LIC enables significant model compression without causing substantial BD‑rate loss. This demonstrates that the model can be reduced in size while maintaining high visual quality. Because the method is post‑training, it requires no retraining, making it ideal for existing production pipelines. The small calibration set also keeps data overhead minimal, which is critical for devices with limited storage. The method also helps reduce bandwidth requirements due to smaller file sizes.

Comparison with Other PTQ Methods

HAMP‑LIC consistently outperforms both fixed‑precision and mixed‑precision PTQ approaches across multiple datasets. Its advantage stems from Hessian‑based sensitivity, a task‑aware refinement that balances distortion and rate‑distortion, and the elimination of cross‑platform encoding‑decoding errors—a common issue in other methods. These results are shown on popular LIC models such as Minnen2018 and Cheng2020. These advantages make HAMP‑LIC an attractive option for service providers seeking to optimize image quality without increasing infrastructure costs.

Conclusion and Future Outlook

HAMP‑LIC represents a significant step toward making LIC viable for edge deployments. By combining Hessian analysis, adaptive precision allocation, and block‑wise calibration, it offers a practical solution to reduce model size and computational load without sacrificing quality. Developers looking to deploy LIC on heterogeneous hardware can adopt HAMP‑LIC within their existing pipelines. Future work may explore automating the calibration process and extending the approach to other generative models.

Gaionix documentation