Zero-Shot Burst Restoration via Diffusion MAP Inference with Poisson-Gaussian Noise Likelihood

Burst image restoration aims to reconstruct a high-quality image from a sequence of low-quality frames. While state-of-the-art supervised methods achieve excellent performance, they rely on extensive paired datasets that may be difficult to acquire in practice and often fail to generalize across different camera sensors or real-world noise levels. Zero-shot diffusion-based methods offer a compelling, training-free alternative. However, they typically struggle on burst imaging due to their assumption of known forward operators. In this paper, we introduce ZSBR: Zero-Shot Burst Restoration, a zero-shot diffusion framework for burst restoration that performs MAP-style inference by combining a pretrained diffusion prior with a physically grounded burst formation model. ZSBR accounts for signal-dependent RAW noise by leveraging a continuous Poisson-Gaussian model. Moreover, unlike common diffusion-based methods that assume a fixed operator, ZSBR refines image alignment parameters online by optimizing per-frame affine matrices within the forward model during sampling. To stabilize the resulting optimization, we integrate an adaptive descent scheme into the sampling loop. Moreover, the method is flexible and allows to achieve a desired distortion-perception tradeoff. Experimental results show that our method provides a flexible and robust solution to burst restoration without requiring any paired training example.