This paper presents a novel diffusion-based auto-bidding framework that uses graph representations to model large-scale auction environments. In such environments, agents must optimize bidding strategies dynamically, balancing key performance indicators (KPIs) while navigating competitive, uncertain, and sparse conditions. To address these challenges, we introduce an approach that combines learnable graph embeddings with a planning-based latent diffusion model (LDM). This model captures the intricate relationships between impression opportunities and multi-agent interactions, enabling more accurate predictions of auto-bidding outcomes. Through reward-alignment techniques, the LDM's posterior is fine-tuned to maximize KPI performance under predefined constraints. Our experiments, conducted in both real-world and synthetic auction environments, show significant improvements in the accuracy of auction outcome forecasts through learnable graph-based embeddings and in bidding performance across a range of common KPIs.