Image-GS: Content-Adaptive Image Representation via 2D Gaussians (github.com)

Researchers have introduced Image-GS, a novel image representation method that reconstructs images using a content-adaptive set of anisotropic, colored 2D Gaussians. Inspired by advances in radiance field rendering, Image-GS leverages a custom differentiable renderer to dynamically allocate and optimize these Gaussian primitives, achieving an efficient balance between visual quality and memory footprint. This approach excels at representing stylized images with non-uniform features and performs well under low-bitrate constraints, marking a significant advancement for graphics and AI-driven image compression. Technically, Image-GS supports hardware-friendly, rapid random access, requiring only 0.3K multiply-accumulate operations (MACs) per pixel decode, making it suitable for real-time applications. It naturally builds a smooth level-of-detail hierarchy through error-guided progressive optimization, allowing flexible quality control and effective rate-distortion trade-offs. The framework supports various bit precision controls for Gaussian parameters, and saliency-guided initialization to better capture image structure. Demonstrated applications include texture compression, semantics-aware image compression, and joint image compression with restoration, underscoring the method’s versatility and potential impact on efficient image processing workflows in AI and graphics contexts.