Reproduce “DLF: Extreme Image Compression with Dual-generative Latent Fusion” 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
Naifu Xue, Zhaoyang Jia, Jiahao Li, Bin Li, Yuan Zhang, Yan Lu
DLF, a novel dual-generative latent fusion framework, achieves state-of-the-art extreme image compression by decomposing latent features into semantic and detail components, leading to superior perceptual quality and reconstruction fidelity at bitrates below 0.01 bpp. It demonstrates BD-rate savings of 43.05% for LPIPS and 67.82% for DISTS compared to MS-ILLM on the Kodak dataset, while offering significantly faster decoding than diffusion models.
Key insights
- Decomposing latent representations into distinct semantic and detail components allows for a more flexible and adaptive compression strategy, addressing both dataset-level commonalities and unique individual object features.
- An intelligent cross-branch interaction mechanism is essential for minimizing redundancy between parallel information streams and enabling complementary learning, leading to optimal bit allocation.
- A hybrid quantization strategy, combining Vector Quantization for semantic tokens and spatially adaptive Scalar Quantization for details, proves critical for achieving superior fidelity by accommodating diverse detail characteristics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.