Show the Signal, Hide the Noise: Spectral Forcing for Pixel-Space Diffusion

This repository contains the weights for Spectral Forcing.

Introduction

Spectral Forcing (SF) is a parameter-free, time-conditional low-pass operator that makes the coarse-to-fine structure of diffusion explicit at the input of a pixel-space model. It is based on the observation that under rectified-flow diffusion and natural-image power-law spectra, a moving frequency front separates a signal-bearing low-frequency region from a noise-dominated high-frequency region at each timestep, and that a standard denoiser wastes capacity outside this region by reproducing closed-form maps rather than modeling the data. SF imposes this boundary directly via a 2D-DCT mask whose cutoff expands with the diffusion time and becomes the identity at the data endpoint, showing the network the signal and hiding the noise.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support