LotteryCodec: Searching the Implicit Representation in a Random Network for Low-Complexity Image Compression

We introduce and validate the lottery codec hypothesis, which states that untrained subnetworks within randomly initialized networks can serve as synthesis networks for overfitted image compression, achieving rate-distortion (RD) performance comparable to trained networks. This hypothesis leads to a new paradigm for image compression by encoding image statistics into the network substructure. Building on this hypothesis, we propose LotteryCodec, which overfits a binary mask to an individual image, leveraging an over-parameterized and randomly initialized network shared by the encoder and the decoder. To address over-parameterization challenges and streamline subnetwork search, we develop a rewind modulation mechanism that improves the RD performance. LotteryCodec outperforms VTM and sets a new state-of-the-art in single-image compression. LotteryCodec also enables adaptive decoding complexity through adjustable mask ratios, offering flexible compression solutions for diverse device constraints and application requirements.

Image compression helps reduce the size of images for storage and transmission, but high performance often requires training large neural networks. In this work, we explore an alternative approach: utilizing untrained parts of randomly initialized networks to compress images effectively. We call this idea the lottery codec hypothesis—suggesting that good subnetworks already exist within randomly created networks, ready to be “found” rather than trained. Building on this, we introduce a new compression approach named LotteryCodec, which finds a binary mask to search subnetworks tailored to each image. LotteryCodec not only surpasses traditional compression tools in terms of quality and size but also enables users to flexibly balance quality and decoding speed, depending on the device, making it ideal for both high-performance and resource-limited devices.