PerCoV2: Advanced Image Compression for Extremely Low Bitrates

In today's world, where images and videos constitute an increasingly large part of our digital communication, the efficient compression of this data plays a crucial role. Especially in areas with limited bandwidth or storage capacity, such as mobile applications or satellite communication, reducing file size while maintaining image quality is of great importance. A promising approach in this field is perceptual image compression, which takes into account the characteristics of human perception to increase compression efficiency. A new method called PerCoV2 promises significant advances in this area.

PerCoV2 builds on the foundations of the system developed by Careil et al. and extends it for the Stable Diffusion 3 environment. The innovative aspect of PerCoV2 lies in the explicit modeling of the discrete hyper-latent image distribution. This allows for more efficient entropy coding, which in turn leads to smaller file sizes. In contrast to traditional compression methods, which focus on pixel-perfect reproduction, perceptual compression considers how the human eye perceives images. This allows information that is less relevant to human perception to be compressed more effectively or even removed without noticeably affecting the subjectively perceived image quality.

A central component of PerCoV2 is the use of autoregressive methods for entropy modeling. The developers have conducted a comprehensive comparative study with current methods such as VAR and MaskGIT and evaluated the results on the extensive MSCOCO-30k benchmark. The results show that PerCoV2 achieves higher image fidelity at even lower bitrates compared to previous approaches, while maintaining competitive perceived quality. Another advantage of PerCoV2 is the hybrid generation mode, which enables additional bitrate savings.

A notable aspect of PerCoV2 is the exclusive use of publicly available components. This promotes transparency and reproducibility of the research results and allows the community to build upon and further develop the code and trained models. The publication of the code and trained models on GitHub underscores this open-source approach and allows other researchers and developers to test, adapt, and improve PerCoV2. This contributes to the advancement of image compression technology and accelerates the development of innovative solutions for future applications.

The implications of PerCoV2 are far-reaching and affect various areas where efficient image compression plays a crucial role. From optimizing image transmission in mobile networks to archiving large image databases, PerCoV2 offers the potential to fundamentally change the storage and transmission of image data. The combination of high compression efficiency and good perceived image quality makes PerCoV2 a promising approach for future applications in image processing and communication.

Bibliography: https://arxiv.org/abs/2503.09368 https://arxiv.org/html/2503.09368v1 http://paperreading.club/page?id=291525 https://huggingface.co/papers?q=MSCOCO-30k https://github.com/wangkai930418/awesome-diffusion-categorized https://openreview.net/forum?id=ktdETU9JBg https://openaccess.thecvf.com/content/CVPR2024/papers/Jia_Generative_Latent_Coding_for_Ultra-Low_Bitrate_Image_Compression_CVPR_2024_paper.pdf https://github.com/ppingzhang/Deep-Learning-Based-Image-Compression https://openaccess.thecvf.com/content_CVPRW_2020/papers/w7/Akutsu_Ultra_Low_Bitrate_Learned_Image_Compression_by_Selective_Detail_Decoding_CVPRW_2020_paper.pdf