The recent developments in the field of data compression and video representation have shown significant advancements, particularly in the application of machine learning and neural networks to enhance efficiency and quality. A notable trend is the exploration of diffusion models for both image and video compression, leveraging their generative capabilities to achieve competitive compression rates and perceptual quality. These models are being adapted to handle temporal information in videos, introducing innovative strategies like Temporal Diffusion Information Reuse to improve inference efficiency. Another emerging direction is the use of Gaussian splats for video representation, offering a novel approach to exploit temporal redundancy and manage dynamic content effectively. Furthermore, the integration of reinforcement learning in video compression algorithms marks a shift towards task-aware optimization, where compression is tailored to enhance the performance of downstream AI tasks rather than solely focusing on human perceptual quality. This approach demonstrates the potential for more efficient and application-specific video encoding methods. Additionally, advancements in nonlinear reduction strategies for data compression are being explored, with a focus on creating a unified framework to compare and understand the effectiveness of various methods across different problem domains. These developments collectively indicate a move towards more intelligent, efficient, and adaptable compression technologies that can cater to a wide range of applications and requirements.

Noteworthy Papers

• Lossy Compression with Pretrained Diffusion Models: Introduces a complete implementation of the DiffC algorithm using Stable Diffusion, achieving competitive compression at ultra-low bitrates without additional training.
• GaussianVideo: Efficient Video Representation Through 2D Gaussian Splatting: Proposes a novel video representation method using 2D Gaussian splats, achieving high compression efficiency and rendering speed.
• RL-RC-DoT: A Block-level RL agent for Task-Aware Video Compression: Develops a reinforcement learning-based approach for optimizing video compression for downstream AI tasks, improving task performance at given bit rates.
• Diffusion-based Perceptual Neural Video Compression with Temporal Diffusion Information Reuse: Presents a diffusion-based framework for video compression that incorporates temporal context and introduces strategies to enhance inference efficiency and handle variable bitrates.