Content Adaptive Switchable Hyperprior Networks for Learned Image Compression

Deniffel S, Seiler J, Kaup A (2026)

Publication Language: English

Publication Status: Published

Publication Type: Conference contribution, Conference Contribution

Future Publication Type: Conference contribution

Publication year: 2026

Publisher: IEEE

Conference Proceedings Title: 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

Event location: Barcelona ES

DOI: 10.1109/ICASSP55912.2026.11461986

Abstract

With the rapid advancement of learned image coders, entropy
models have proven to be crucial for achieving state of the
art efficiency. These models estimate the latent space statis-
tics to enable more efficient arithmetic compression. Key to
almost any entropy model are the hyperprior networks. Usu-
ally, those are trained once and utilized for all images during
inference. This poses a suboptimal solution due to the diverse
image content typically found in compression tasks. We pro-
pose a coder-agnostic solution to this challenge which we call
Content Adaptive Switchable Hyperprior networks (CASH).
We employ the best of multiple specialized hyperprior net-
works on an individual image basis to achieve a lower bitrate.
To show the practical applicability of our technique, we in-
tegrate it with two learned image coders and introduce two
additional variants, leveraging a single shared hyperprior de-
coder and sharing certain layers between hyperprior encoders.
Our new architecture achieves a BD-rate of up to −3.05%.
With our additional variants, the increase in number of param-
eters is limited to just 0.49% and only to the encoder, enabling
wide applicability.

Authors with CRIS profile

Simon Deniffel Lehrstuhl für Multimediakommunikation und Signalverarbeitung (LMS) Jürgen Seiler Lehrstuhl für Multimediakommunikation und Signalverarbeitung (LMS) Andre Kaup Lehrstuhl für Multimediakommunikation und Signalverarbeitung (LMS)

How to cite

APA:

Deniffel, S., Seiler, J., & Kaup, A. (2026). Content Adaptive Switchable Hyperprior Networks for Learned Image Compression. In 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). Barcelona, ES: IEEE.

MLA:

Deniffel, Simon, Jürgen Seiler, and André Kaup. "Content Adaptive Switchable Hyperprior Networks for Learned Image Compression." Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Barcelona IEEE, 2026.

BibTeX: Download