Image-GS represents images using content-adaptive 2D Gaussians that are progressively optimized through a custom differentiable renderer. It adaptively allocates anisotropic, colored Gaussians based on local image complexity, achieving efficient compression with hardware-friendly random access.
Enables flexible, real-time image representation for graphics applications with only 0.3K MACs per pixel decode. Supports texture compression, semantics-aware compression, and joint compression-restoration while maintaining visual quality in low-bitrate scenarios.
