Imperial College London
ICML 2025 (spotlight)
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-parametrization challenges and streamline subnetwork search, we develop a rewind modulation mechanism that improves the RD performance. LotteryCodec 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.
Let \( d \) denote a distortion function and \( H \) the entropy function. For any overfitted image codec \( g_{\mathbf{W}}(\mathbf{z}) \), there exists an over-parameterized and randomly initialized network \( g_{\mathbf{W'}} \) with \( |\mathbf{W'}| > |\mathbf{W}| \) and a pair \( (\mathbf{\tau'}, \mathbf{z'}) \) as the ‘winning tickets’, such that \( d(\mathbf{S}, \mathbf{S}') \le d(\mathbf{S}, \mathbf{S}^*) \) and \( H(\mathbf{\hat{z}}') = H(\mathbf{\hat{z}}) \).
Fig. 1: From AE-based neural codec to LotteryCodec.
Fig. 2: (a) RD curve and BD-rate under different over-parameterization, (b) BD-rate vs. mask ratios, (c)-(d) BD-rate across width/depth variations, where \( (N_t, d) \) refers to \( N_t \) hidden layers and \( d \) hidden dimensions.
Fig. 3: The source image is encoded into a binary mask and latent modulations. During decoding, the receiver initializes a random network and uses a modulated subnetwork to reconstruct the source.
Fig. 4: SuperMask maps coordinates to RGB values by identifying subnetworks within a randomly initialized network, guided by modulations generated by the ModNet using latent input.
(a) RD curve and BD rate on Kodak dataset. (b) RD curve and BD rate on CLIC2020 dataset. (c) BD-rate and decoding complexity across different mask ratios on Kodak dataset.
@article{LotteryCodec,
title={LotteryCodec: Searching the Implicit Representation in a Random Network for Low-Complexity Image Compression},
author={Haotian Wu, Gongpu Chen, Pier Luigi Dragotti, and Deniz Gündüz},
journal={International Conference on Machine Learning (ICML) 2025},
year={2025}
}