Sampling-Efficient Unconditional Pure-Deblurring Diffusion Models via Noise-Augmented Generation

Diffusion models have proven to be powerful image priors, supporting unconditional generation and a range of conditional tasks such as image restoration and text-guided synthesis. They learn to reverse a progressive noising process, capturing regularities of the data distribution. Nevertheless, diffusion models retain drawbacks that include slow sampling, high computational cost, and nontrivial architectural complexity.Recent work has explored replacing noise-based corruption with more structured degradations, most commonly progressive blurring, which aligns with the multi-scale structure of images and has been shown to support high-quality generation. However, sampling requires introducing stochasticity into the otherwise deterministic blurring process, typically by adding Gaussian noise, which causes the reverse dynamics to remain, in effect, a denoising process.Our blurring-based approach applies Gaussian noise, solely as a data augmentation, symmetrically to both inputs and targets, allowing the network to learn a pure and simple deblurring operation, with the noise used only to regularize training and enable stochastic sampling. This separation between blur and noise yields a lightweight generative process that produces high-quality samples using only a small number of sampling steps, without requiring dedicated fast-sampling mechanisms. Empirical results confirm the effectiveness of the proposed approach and its practical viability for generative modeling.