Colored Noise Diffusion Sampling

Researchers introduced Colored Noise Sampling, a training-free diffusion sampler that injects noise by frequency band and timestep instead of using uniform white noise.

The method targets a model’s spectral bias, matching noise to unresolved frequency components during inference.

Across multiple architectures and solvers, it improved FID over standard ODE and SDE sampling, including on ImageNet-256.

The gains held with and without classifier-free guidance, suggesting a plug-and-play way to improve image generation without retraining.