DiFA: Inference-Time Forward-Process Alignment for Diffusion Models
Merged summary
TL;DR - DiFA is a training-free diffusion inference framework that treats iterative denoising predictions as correlated observations in a sequential state-estimation process. It improves generative fidelity by building a forward-process-aligned consensus while preserving fine details.
- Aggregates historical reverse-trajectory predictions based on structural consistency and noise-level compatibility.
- Draws on Kalman filtering rather than treating inference solely as numerical integration.
- Uses adaptive deviation guidance to counter temporal consensus over-smoothing.
- Improves FID, IS, and FD-DINOv2 metrics on CIFAR-10 and ImageNet.
Sources (1)
DiFA: Inference-Time Forward-Process Alignment for Diffusion Models
TL;DR - DiFA is a training-free diffusion inference framework that treats iterative denoising predictions as correlated observations in a sequential state-estimation process. It improves generative fidelity by building a forward-process-aligned consensus while preserving fine details.
- Aggregates historical reverse-trajectory predictions based on structural consistency and noise-level compatibility.
- Draws on Kalman filtering rather than treating inference solely as numerical integration.
- Uses adaptive deviation guidance to counter temporal consensus over-smoothing.
- Improves FID, IS, and FD-DINOv2 metrics on CIFAR-10 and ImageNet.