The paper introduces SWAG, a novel method for 3D scene reconstruction from unconstrained photo collections, extending the 3D Gaussian Splatting (3DGS) technique. 3DGS, while efficient and visually appealing, struggles with varying lighting conditions and transient objects in *in-the-wild* datasets. SWAG addresses these limitations by incorporating appearance modeling and transient object handling. Specifically, it uses a learned embedding space to capture photometric variations in the rendered images and introduces a mechanism to train transient Gaussians for handling occluders. Experiments on the Phototourism dataset and the NeRF-OSR benchmark demonstrate that SWAG achieves state-of-the-art performance in novel view synthesis, with improved efficiency and real-time rendering capabilities. The method effectively models varying appearances and removes transient objects, making it suitable for diverse and challenging real-world scenes.The paper introduces SWAG, a novel method for 3D scene reconstruction from unconstrained photo collections, extending the 3D Gaussian Splatting (3DGS) technique. 3DGS, while efficient and visually appealing, struggles with varying lighting conditions and transient objects in *in-the-wild* datasets. SWAG addresses these limitations by incorporating appearance modeling and transient object handling. Specifically, it uses a learned embedding space to capture photometric variations in the rendered images and introduces a mechanism to train transient Gaussians for handling occluders. Experiments on the Phototourism dataset and the NeRF-OSR benchmark demonstrate that SWAG achieves state-of-the-art performance in novel view synthesis, with improved efficiency and real-time rendering capabilities. The method effectively models varying appearances and removes transient objects, making it suitable for diverse and challenging real-world scenes.