One-Step Diffusion for Perceptual Image Compression

TL;DR

This paper tackles the latency challenge of diffusion-based perceptual image compression by proposing OSdiff, which decodes in a single denoising step. It combines a frozen Stable Diffusion encoder/decoder with lightweight learnable modules and a discriminator that operates in a latent feature space to boost realism. The approach introduces a one-step sampling mechanism and a multi-term objective that jointly optimizes diffusion accuracy, coding rate, and latent fidelity, yielding competitive rate-distortion-perception while achieving about 46x faster decoding than multi-step diffusion methods. Experiments on standard datasets demonstrate strong perceptual metrics (LPIPS, DISTS) and favorable inference speed, supporting practical deployment. The work provides open-source code and models, facilitating adoption and further research in fast diffusion-based compression.

Abstract

Diffusion-based image compression methods have achieved notable progress, delivering high perceptual quality at low bitrates. However, their practical deployment is hindered by significant inference latency and heavy computational overhead, primarily due to the large number of denoising steps required during decoding. To address this problem, we propose a diffusion-based image compression method that requires only a single-step diffusion process, significantly improving inference speed. To enhance the perceptual quality of reconstructed images, we introduce a discriminator that operates on compact feature representations instead of raw pixels, leveraging the fact that features better capture high-level texture and structural details. Experimental results show that our method delivers comparable compression performance while offering a 46$\times$ faster inference speed compared to recent diffusion-based approaches. The source code and models are available at https://github.com/cheesejiang/OSDiff.

Figures (6)

• Figure 1: Qualitative comparisons of different methods on test datasets.
• Figure 2: The framework of our OSDiff, which is composed of the frozen encoder–decoder pair ($E_{SD}$, $D_{SD}$), the modules $G_a$ and $G_s$, the discriminator $D$, and the denoising network $\epsilon_{\theta}$ with control module ($ctr$ for short). During training, only the modules $G_a$, $G_s$, $ctr$, and the discriminator $D$ are optimized through the loss function, while the parameters of other components remain frozen.
• Figure 3: The distribution of generated features and real features in the specific lantent feature space.
• Figure 4: Quantitative comparisons with the state-of-the-art method on test datasets.
• Figure 5: Qualitative comparisons of different methods on test datasets.