Not All Noise Is Harmful: Towards Perception Aware and Controllable RAW Image Joint Denoising and Demosaicing

RAW image denoising and demosaicing play a critical role in the early stages of the image signal processing (ISP) pipeline. Previous methods handle them sequentially, which can introduce error propagation. Unified restoration, such as joint denoising and demosaicing (JDD), has become a widely adopted paradigm and achieves impressive performance. Nonetheless, under constraints of model capacity and low-light conditions, these methods tend to apply overly aggressive denoising, causing notable smearing effects and the erosion of fine textures. Researchers attempt to moderate the denoising strength to better preserve details, but this often results in noticeable artifacts. These observations suggest a key insight: Not All Noise is Harmful—retaining an appropriate amount of noise can be preferable to over-smoothing and is often aligned with user aesthetics. Moreover, most existing methods are mainly trained on paired clean–degraded data, yielding fixed, hard-to-control denoising behavior that fails to meet evolving user demands. In this paper, we propose $\textbf{JDD-PRO}$, a $\textbf{P}$erception-awa$\textbf{R}$e and contr$\textbf{O}$llable framework for RAW image JDD. Built on a controllable JDD training scheme and a HVS-inspired perceptual adapter, our method allows users to tune the denoising strength at inference to strike their desired balance between noise removal and detail preservation. Comprehensive experiments on real-world datasets validate the effectiveness and robustness of our proposed method.