Colored Noise Diffusion Sampling

Researchers introduced Colored Noise Sampling, a plug-and-play diffusion sampler that injects frequency-aware noise during inference.

The method frames diffusion sampling as frequency-decoupled energy transfer, replacing uniform white noise with colored noise.

Across models such as SiT, JiT, and FLUX on ImageNet-256, it improves image quality and lowers FID without retraining.

Because it is training-free and model-agnostic, the approach offers an easy way to boost diffusion sampling performance.