GaussianImage is a novel paradigm for image representation and compression which is based on 2D Gaussian Splatting. This approach aims to address the limitations of traditional image representation methods and implicit neural representations (INRs) by leveraging the efficiency and expressiveness of 2D Gaussian representations. The key innovation lies in the adoption of 2D Gaussians instead of 3D Gaussians, resulting in a more compact and expressive representation with significant storage savings. Additionally, a unique rasterization algorithm is proposed, which replaces depth-based Gaussian sorting and alpha blending with an accumulated summation process. This approach not only improves fitting performance but also accelerates training and inference speed. The transfer of the 2D Gaussian representation into a practical image codec employs a two-step compression strategy involving attribute quantization-aware fine-tuning and encoding. The use of vector quantization and partial bits-back coding further enhances the compression performance of the codec.

The proposed GaussianImage approach offers several notable contributions. Firstly, it achieves high representation performance with swift training, minimal GPU memory overhead, and remarkably fast rendering speed. Secondly, the development of a low-complexity neural image codec using vector quantization, along with the optional use of partial bits-back coding, demonstrates the practical applicability of the approach in image compression. The experimental results showcase the remarkable training and inference acceleration, reduced GPU memory usage, and competitive compression performance of the GaussianImage approach compared to existing INR methods and state-of-the-art image codecs.