Research map / Image generation / Image restoration
Computationally-Efficient Neural Image Compression with Shallow Decoders
Yibo Yang, Stephan Mandt
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SGA iterative encoding to the two-layer shallow synthesis (ablation) | 2-layer synthesis · Kodak | BD rate savings vs BPG (%): -5.19 -> 4.67 | Improved |
| use two-layer nonlinear synthesis instead of JPEG-like linear synthesis (ablation) | shallow decoder · Kodak | BD rate savings vs BPG (%): -20.95 -> -5.19 | Improved |
| increase JPEG-like synthesis kernel size to k=26 (ablation) | JPEG-like synthesis · Kodak | bit-rate vs linear CNN synthesis: within 1% of linear CNN with 94% fewer FLOPs | Improved |
| use minimal kernel size k=s=16 in JPEG-like synthesis (ablation) | JPEG-like synthesis · Kodak | visual artifacts | Hurt |
| replace ELIC analysis transform with Mean-Scale Hyperprior analysis (ablation) | 2-layer synthesis · Kodak | bit-rate vs BPG: over 6% worse | Hurt |
| add residual connection to the two-layer synthesis (ablation) | 2-layer synthesis · Kodak | BD rate savings | Improved |
| use inverse GDN activation and larger output kernel size (ablation) | 2-layer synthesis · Kodak | BD rate savings | Improved |
| evaluate shallow synthesis on perceptual metrics MS-SSIM and LPIPS | 2-layer synthesis + SGA · Kodak, Tecnick, CLIC | BD-rate vs Hyperprior: 8% to 12% worse | Hurt |
| pair a shallow decoder with a stronger encoder and iterative encoding | Mean-Scale Hyperprior · Kodak | decoding complexity: 108.97 -> 20.52 KMAC per pixel | Improved |
Closest papers by meaning
- Full Resolution Image Compression With Recurrent Neural Networks (CVPR 2017 · similarity 0.91)
- End-to-end Optimized Image Compression (ICLR 2017 · similarity 0.91)
- Neural Image Compression: Generalization, Robustness, and Spectral Biases (NeurIPS 2023 · similarity 0.91)
- C3: High-Performance and Low-Complexity Neural Compression from a Single Image or Video (CVPR 2024 · similarity 0.89)
- Controllable Distortion-Perception Tradeoff Through Latent Diffusion for Neural Image Compression (AAAI 2025 · similarity 0.89)
- Robustly overfitting latents for flexible neural image compression (NeurIPS 2024 · similarity 0.89)
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.
