Recent advances in scene reconstruction have pushed toward highly realistic modeling of autonomous driving (AD) environments using 3D Gaussian splatting. However, the resulting reconstructions remain closely tied to the original observations and struggle to support photorealistic synthesis of significantly altered or novel driving scenarios. This work introduces MADrive, a memory-augmented reconstruction framework designed to extend the capabilities of existing scene reconstruction methods by replacing observed vehicles with visually similar 3D assets retrieved from a large-scale external memory bank. Specifically, we release MAD-Cars, a curated dataset of ~70K 360° car videos captured in the wild and present a retrieval module that finds the most similar car instances in the memory bank, reconstructs the corresponding 3D assets from video, and integrates them into the target scene through orientation alignment and relighting. The resulting replacements provide complete multi-view representations of vehicles in the scene, enabling photorealistic synthesis of substantially altered configurations, as demonstrated in our experiments.

Given an input frame sequence, our retrieval scheme finds similar vehicles in an external database (Left). The 3D reconstruction pipeline then produces detailed vehicle models from the retrieved videos. The vehicles are represented with relightable 2D Gaussian splats. Opacity masks are used to remove background splats. The model geometry is regularized with external normals maps. (Middle) Finally, the reconstructed vehicles are adapted to the scene’s lighting conditions and composed with the background to produce the overall scene representation (Right).