G-CUT3R presents a novel feed-forward approach for guided 3D scene reconstruction that enhances the CUT3R model by incorporating prior information. Unlike existing feed-forward approaches that rely solely on input images, it leverages auxiliary data commonly found in real-world scenarios, such as depth, camera calibration, and camera position. We propose a lightweight modification to CUT3R that integrates dedicated encoders for each modality and fuses them with RGB image tokens via zero convolution. This flexible design allows for seamless integration of any combination of prior information during inference. Evaluations on multiple benchmarks and multi-view tasks, including 3D reconstruction, demonstrate that the proposed approach achieves significant performance improvements, effectively utilizes available prior information, and maintains compatibility with diverse input modalities.
