Reproduce “CoD-Lite: Real-Time Diffusion-Based Generative Image Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhaoyang Jia, Naifu Xue, Zihan Zheng, Jiahao Li, Bin Li, Xiaoyi Zhang, Zongyu Guo, Yuan Zhang and 2 more
CoD-Lite presents a diffusion-based image codec that achieves real-time encoding (60 FPS) and decoding (42 FPS) for 1080p images on an NVIDIA A100 GPU, offering a substantial speed improvement over previous diffusion-based compression methods. This lightweight convolutional framework maintains high perceptual fidelity and operates across a wide bitrate range (0.0039 to 0.5 bpp), addressing a key limitation in practical generative image compression.
Key insights
- Compression-oriented pre-training with information-dense conditions is uniquely effective for lightweight diffusion models, providing significant performance gains compared to generic generation-oriented pre-training.
- Global attention mechanisms, while crucial for general image generation, are largely redundant in compression-oriented diffusion models; local convolutional operators, combined with distillation, can achieve comparable perceptual quality at significantly faster speeds.
- A lightweight convolutional diffusion backbone, when guided by distillation from a larger teacher model and further refined with adversarial training, can reconcile real-time performance with high perceptual quality in generative compression.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.