RESEARCH · RESEARCH · #1632
Energy-conditioned noise schedule and spectral whitening for transform-domain diffusion (arXiv:2610.07206v1)
This arXiv preprint introduces a global spectral whitening step combined with an energy-conditioned noise allocation for transform-domain (spectral) diffusion models. The proposed forward process preserves Gaussian transitions and is compatible with standard DDPM/DDIM sampling without changing model architecture; on CIFAR-10 a compact DCTdiff U-Net variant reduced FID from 142.48 to 100.45.
KEY POINTS
- This arXiv preprint introduces a global spectral whitening step combined with an energy-conditioned noise allocation for transform-domain (spectral) diffusion models.
- The proposed forward process preserves Gaussian transitions and is compatible with standard DDPM/DDIM sampling without changing model architecture; on CIFAR-10 a compact DCTdiff U-Net variant reduced FID from 142.48 to 100.45.
- Aligning the forward diffusion noise schedule with per-coefficient energy and applying spectral whitening improves sample quality for spectral diffusion methods without requiring architectural changes.
WHY IT MATTERS
Aligning the forward diffusion noise schedule with per-coefficient energy and applying spectral whitening improves sample quality for spectral diffusion methods without requiring architectural changes.