DiFA: Inference-Time Forward-Process Alignment for Diffusion Models

DiFA is a training-free inference framework that treats successive data predictions during diffusion sampling as correlated observations in a sequential state-estimation problem. Inspired by Kalman filtering, it forms a forward-process-aligned temporal consensus using structural consistency and noise-level compatibility. A deviation-guidance mechanism is added to preserve residual details that consensus aggregation might smooth away. The authors report improvements across FID, IS, and FD-DINOv2 on CIFAR-10 and ImageNet, although the abstract does not provide numerical gains, runtime overhead, sampler coverage, or ablation details.