Advancing 2D Image Modelling through Gaussian Splatting Techniques

We present an adaptation of 3D Gaussian splatting for 2D grayscale image modeling using gradient descent optimization. Images are approximated as a sum of localized 2D Gaussian primitives, refined through differentiable loss functions. To benchmark our method, we compare it against a brute-force baseline, showing that while gradient descent is efficient, explicit solutions should be preferred when computable. Our 2D focus simplifies experimentation and enables modeling of standalone image structures. Results demonstrate effective approximation with potential for image and video representations.