3D Gaussian Splatting has emerged as a promising method for modeling 3D scenes using Gaussian mixtures. Existing optimization methods rely on backpropagation through differentiable rendering pipelines, but suffer from the critical forgetting problem when processing continuous data streams. In this paper, we propose Variational Bayes Gaussian Splatting (VBGS), a novel method that constructs Gaussian splat training as variational inference for model parameters. Leveraging the conjugate nature of multivariate Gaussians, we derive a closed-form variational update rule, enabling efficient updates from partial, sequential observations without a replay buffer. Experimental results show that VBGS not only achieves state-of-the-art performance on static datasets, but also significantly improves performance in these settings by enabling continuous learning from sequentially streaming 2D and 3D data.
