DUNE: Training-free refinement boosts diffusion model fidelity and cuts hallucinations

A new paper from Haksoo Lim and colleagues (KAIST) presents DUNE, a training-free refinement framework for diffusion models. The key insight is that abrupt, early-stage fluctuations in deep latent representations are strongly associated with image artifacts. DUNE detects these deviations using a shared EMA-based criterion that monitors internal latents at low-noise steps. Once problematic entries are identified, the framework applies backbone-specific suppression to stabilize the generation process. Although derived from U-Net architectures, the same detect-suppress principle extends naturally to Transformer-based diffusion models by acting on latents of deep self-attention blocks.

Extensive experiments across multiple backbones show that DUNE improves fidelity while reducing hallucinations, all without any additional training or fine-tuning. The method provides a systematic, phase-aware analysis of diffusion components, offering new insight into where and when diffusion backbones should be controlled. Accepted at the European Conference on Computer Vision (ECCV) 2026, this work is a practical tool for developers using state-of-the-art image generation models, enabling cleaner outputs with minimal overhead.