Research map / Image generation / Image restoration
LotteryCodec: Searching the Implicit Representation in a Random Network for Low-Complexity Image Compression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LotteryCodec, an overfitted image codec that transmits only a learned binary mask over a frozen random-init network plus latent modulations, beating VTM-19.1 in BD-rate with low decoding cost.
Problem
Overfitted INR-based image codecs lag behind classical codecs like VTM in rate-distortion performance, and increasing fidelity requires larger networks that raise the bitrate; fixed architectures also limit adaptivity.
Approach
Test the lottery codec hypothesis: an untrained subnetwork of a random over-parameterized network can serve as the synthesis network. Encode each image as a learned binary mask (supermask) plus latent modulations, with a ModNet and a rewind modulation mechanism that concatenates modulation vectors to each layer to ease subnetwork search.</может
Key results
- On Kodak, LotteryCodec achieves a BD-rate of -6.4% over C3 and reaches VTM-19.1 level performance, outperforming VTM on both Kodak and CLIC2020.
- On CLIC2020, the gain over VTM-19.1 reaches up to -9.79% BD-rate.
- Compared with MLIC+, LotteryCodec needs about two orders of magnitude fewer decoding MACs, with a +2.33% BD-rate gap relative to VTM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LotteryCodec instead of C3 | overfitted image compression · Kodak | BD-rate: -6.4% | Helped |
| use LotteryCodec instead of C3-adapt | overfitted image compression · Kodak | BD-rate: -3.73% | Helped |
| use LotteryCodec instead of VTM-19.1 | image compression · CLIC2020 | BD-rate: -9.79% | Helped |
| use LotteryCodec instead of C3 | overfitted image compression · CLIC2020 | BD-rate: -5.69% | Helped |
| use LotteryCodec instead of C3-adpt | overfitted image compression · CLIC2020 | BD-rate: -2.90% | Helped |
| use LotteryCodec instead of COOL-CHIC v2 | overfitted image compression · CLIC2020 | BD-rate: -25.57% | Helped |
| remove SuperMask (ablation) | LotteryCodec · Kodak | BD-rate: +12.45% | Hurt |
| remove ModNet and use random network (4,32) (ablation) | LotteryCodec · Kodak | BD-rate: +13.02% | Hurt |
| remove ModNet and use random network (4,48) (ablation) | LotteryCodec · Kodak | BD-rate: +11.98% | Hurt |
| remove ModNet and use random network (4,64) (ablation) | LotteryCodec · Kodak | BD-rate: +14.99% | Hurt |
| replace Fourier initialization with Kaiming normal initialization (ablation) | LotteryCodec · First 10 Kodak images | BD rate: +1.55% | Hurt |
| replace Fourier initialization with Signed Kaiming constant initialization (ablation) | LotteryCodec · First 10 Kodak images | BD rate: +0.63% | Hurt |
Surprising findings
- Optimal mask ratio is about 50% without rewind modulation but about 20% with it.
Limitations
- High encoding complexity since each image is overfitted, as with all overfitted codecs.
- Theoretical decoding complexity excludes masked operations; practical speed depends on sparsity-aware libraries.
Experimental setup
Datasets: Kodak CLIC2020
Benchmarks: Kodak CLIC2020
Metrics: PSNR MS-SSIM BD-rate MACs per pixel
Baselines: VTM-19.1 HEVC CST EVC MLIC+ C3 COOL-CHIC v2 C3-adapt ELIC
Builds on: lottery ticket hypothesis strong lottery ticket hypothesis C3 COOL-CHIC Fourier reparameterization
Closest papers by meaning
- What Matters in Practical Learned Image Compression (CVPR 2026 · similarity 0.85)
- Real-Time Adaptive Image Compression (ICML 2017 · similarity 0.85)
- Improved Lossy Image Compression With Priming and Spatially Adaptive Bit Rates for Recurrent Networks (CVPR 2018 · similarity 0.84)
- Practical Full Resolution Learned Lossless Image Compression (CVPR 2019 · similarity 0.84)
- Optimal Neural Compressors for the Rate-Distortion-Perception Tradeoff (NeurIPS 2025 · similarity 0.84)
- Towards image compression with perfect realism at ultra-low bitrates (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.