DiFA: Inference-Time Forward-Process Alignment for Diffusion Models
DiFA treats diffusion sampling as state estimation, using past denoising predictions as correlated evidence rather than just integration steps.
The method is training-free and builds a forward-aligned temporal consensus during inference. It weights historical predictions by structural consistency and noise-level compatibility, with a deviation guidance step meant to preserve detail that consensus can smooth away. The authors report gains on CIFAR-10 and ImageNet across FID, IS, and FD-DINOv2. The paper is accepted to ICML 2026. HF Daily Papers' note
The method is training-free and builds a forward-aligned temporal consensus during inference. It weights historical predictions by structural consistency and noise-level compatibility, with a deviation guidance step meant to preserve detail that consensus can smooth away. The authors report gains on CIFAR-10 and ImageNet across FID, IS, and FD-DINOv2. The paper is accepted to ICML 2026. HF Daily Papers' note
score 4