Image-GS: Content-Adaptive Image Representation via 2D Gaussians
a day ago
- #real-time graphics
- #image compression
- #neural rendering
- Neural image representations offer a balance between visual fidelity and memory efficiency.
- Existing methods often use fixed data structures or compute-intensive models, limiting real-time applications.
- Image-GS introduces a content-adaptive image representation using 2D Gaussians.
- It uses a differentiable renderer to adaptively allocate and optimize anisotropic, colored 2D Gaussians.
- Image-GS achieves high visual fidelity and memory efficiency, especially for stylized images with non-uniform features.
- Supports hardware-friendly rapid random access, requiring only 0.3K MACs per pixel.
- Features error-guided progressive optimization for a smooth level-of-detail hierarchy.
- Demonstrated applications include texture compression, semantics-aware compression, and joint image compression and restoration.