Diffusion models have achieved remarkable success across diverse domains, with performance closely related to the denoising backbones that parameterize the score function. This paper presents a systematic, phase-aware analysis of diffusion components, showing that abrupt early-stage fluctuations in deep latents are strongly associated with artifacts.
Guided by these findings, we introduce DUNE (Diffusion Unified Network refiNEr), a training-free refinement framework that detects abrupt deviations in deep low-noise internal latents using a shared EMA-based criterion, applying backbone-specific suppression to the detector-selected entries.
Although derived from U-Net, the same detect-suppress principle extends naturally to Transformer-based diffusion models by acting on the latents of deep self-attention blocks. Extensive experiments across multiple backbones indicate that DUNE improves fidelity while reducing hallucinations, offering new insight into where and when diffusion backbones should be controlled.
Blogger's Review: DUNE showcases the importance of reducing artifacts in diffusion models through systematic analysis and an innovative training-free framework. Its applicability extends beyond U-Net to Transformer architectures, demonstrating significant flexibility and potential for broader applications.