​
Overview

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.

​
Why it matters

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.

​
Key trade-offs / limitations

  • Linear scaling with number of Gaussians affects memory for complex images
  • Progressive optimization requires multiple training iterations
  • Performance depends on effective gradient-based initialization
  • Limited to 2D image representation (no temporal dynamics yet)
Link arXiv:2407.01866