GLACE: Global Local Accelerated Coordinate Encoding is a novel method for scene coordinate regression (SCR) that enables accurate localization in large-scale scenes using a single, small-sized network without relying on 3D model supervision. The method addresses the challenge of implicit triangulation in large-scale scenes by introducing the concept of co-visibility through global encodings. GLACE integrates pre-trained global and local encodings, enabling the network to scale effectively to large scenes. A novel feature diffusion technique is proposed to implicitly group reprojection constraints with co-visibility, avoiding overfitting to trivial solutions. Additionally, a positional decoder is introduced to more effectively parameterize output positions for large-scale scenes. GLACE achieves state-of-the-art results on large-scale scenes with a low-map-size model, outperforming existing methods such as ACE and Poker on the Cambridge landmarks dataset. The method is efficient, requiring only a single model and avoiding the need for ensembles or 3D model supervision. The approach is validated on multiple datasets, including 7 Scenes, 12 Scenes, Cambridge Landmarks, and Aachen Day, demonstrating its effectiveness in both indoor and outdoor environments. The method's ability to handle large-scale scenes with a single network and without 3D model supervision makes it a promising solution for visual localization tasks.GLACE: Global Local Accelerated Coordinate Encoding is a novel method for scene coordinate regression (SCR) that enables accurate localization in large-scale scenes using a single, small-sized network without relying on 3D model supervision. The method addresses the challenge of implicit triangulation in large-scale scenes by introducing the concept of co-visibility through global encodings. GLACE integrates pre-trained global and local encodings, enabling the network to scale effectively to large scenes. A novel feature diffusion technique is proposed to implicitly group reprojection constraints with co-visibility, avoiding overfitting to trivial solutions. Additionally, a positional decoder is introduced to more effectively parameterize output positions for large-scale scenes. GLACE achieves state-of-the-art results on large-scale scenes with a low-map-size model, outperforming existing methods such as ACE and Poker on the Cambridge landmarks dataset. The method is efficient, requiring only a single model and avoiding the need for ensembles or 3D model supervision. The approach is validated on multiple datasets, including 7 Scenes, 12 Scenes, Cambridge Landmarks, and Aachen Day, demonstrating its effectiveness in both indoor and outdoor environments. The method's ability to handle large-scale scenes with a single network and without 3D model supervision makes it a promising solution for visual localization tasks.