Quality-Aware Modulation Improves Diffusion Transformer Image Fidelity

Modern text-to-image diffusion models, particularly Diffusion Transformers (DiT), typically modulate the denoising process based on timestep or prompt embeddings. While this conveys the current noise level, it lacks explicit information about the desired output quality, often leading to generated images that are misaligned, inconsistent, or lack high fidelity. To address this, a new paper introduces the Quality Representation Module (QRM). This lightweight transformer module is designed to learn a quality-aware representation directly from existing model inputs. It then produces a set of vectors that adjust the adaptive LayerNorm modulation within the DiT transformer blocks. By injecting this quality-sensitive signal into the denoising parameters, QRM consistently improves image quality. Crucially, it achieves these enhancements without requiring significant modifications to the sampling schedule or the core diffusion backbone. Experimental results, including ablations on training losses and architectures, confirm the consistent visual quality improvements over baseline DiT-based models.