Reproduce “Fast Training-free Perceptual 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
Ziran Zhu, Tongda Xu, Minye Huang, Dailan He, Xingtong Ge, Xinjie Zhang, Ling Li, Yan Wang
This work introduces a fast training-free perceptual image compression algorithm that significantly reduces decoding latency while improving perceptual quality. The method applies a three-step post-processing procedure leveraging pre-trained unconditional diffusion models to enhance images decompressed by any base codec, achieving decoding times of 0.1 seconds to 10 seconds and outperforming existing methods in perceptual quality.
Key insights
- A post-processing method leveraging pre-trained unconditional diffusion models can theoretically guarantee a reduction in KL divergence from the true image distribution, thus improving perceptual quality.
- Different implementations of the denoising step (Consistency Models, ODE/SDE solvers, SDE with Diffusion Posterior Sampling) allow for flexible trade-offs between decoding speed and perceptual quality.
- The approach naturally enables perception-distortion trade-off control by simply varying the initial noise level added to the decoded image.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.