Diffusion models have achieved significant advances in image and audio generation through classifier-free guidance, but guidance scale selection has remained understudied. Fixed scales often fail to generalize across prompts of varying complexity and often lead to oversaturation or weak alignment. This paper addresses this gap by introducing a prompt-aware framework to predict scale-dependent quality and select optimal guidance during inference. Specifically, we build a large-scale synthetic dataset by generating samples at multiple scales and scoring them with reliable evaluation metrics. A lightweight predictor conditioned on semantic embeddings and linguistic complexity estimates a multi-metric quality curve, and a utility function is used to determine the optimal scale via regularization. Experimental results on MSCOCO~2014 and AudioCaps demonstrate consistent improvements over vanilla CFG, improving fidelity, alignment, and perceptual desirability. This study demonstrates that prompt-aware scale selection provides effective, training-free enhancements to pretrained diffusion backbones.
