LiVeAction: a Lightweight, Versatile, and Asymmetric Neural Codec Design for Real-time Operation

Modern sensors produce rich data, but bandwidth and power limits on wearables and remote devices often force trade-offs. Standard codecs like JPEG and MPEG optimize for human perception, not machine tasks, and struggle with non-traditional modalities such as spatial audio arrays, hyperspectral images, and 3D medical scans. General-purpose compression like scalar quantization fails to exploit signal redundancies, while recent generative neural codecs are over-parameterized, data-hungry, and modality-specific. LiVeAction solves this with two innovations: an FFT-like structure that drastically reduces encoder size and depth, and a variance-based rate penalty that removes the need for adversarial or perceptual losses, simplifying training across arbitrary signal types.

LiVeAction delivers rate-distortion performance that surpasses state-of-the-art generative tokenizers, yet operates within the tight compute and memory budgets of low-power sensors. This makes it practical for real-time compression on edge devices like drones, medical wearables, and IoT nodes. The authors have released their code, experiments, and a Python library to accelerate adoption. The work was submitted to DCC 2026 and spans multiple domains: image/video processing, audio, and machine learning. For professionals building edge-AI pipelines, LiVeAction offers a drop-in neural codec that is both high-performance and resource-efficient.