Bridging RGB and RAW: Single-step Deterministic Flow with Homogeneous Representation Alignment

Reconstructing high-fidelity RAW sensor data from processed RGB images is a fundamental yet ill-posed problem, plagued by irreversible information loss and complex non-linear ISP transformations. While generative models offer high-quality reconstruction, they suffer from prohibitive computational costs. Conversely, dominant regression-based methods are fast but susceptible to incoherent observational deviations, often yielding over-smoothed predictions that drift from the authentic signal manifold. To reconcile this trade-off, we propose SHADE, a Single-step Homogeneous Aligned DEterministic flow framework. We validate that, unlike point-to-point regression, the single-step deterministic flow captures global transport trends and enables intrinsic robustness against input perturbations. Furthermore, we introduce Homogeneous Representation Alignment to maximize fidelity. By leveraging a homogeneously initialized student-teacher DINO pair, this mechanism enforces alignment within a shared feature space, significantly amplifying the representational capacity. Extensive experiments demonstrate that SHADE achieves state-of-the-art performance on multiple benchmarks, establishing a new paradigm for accurate and efficient sensor data reconstruction.

Modern cameras and smartphones do not usually save images exactly as the sensor sees them. Instead, they apply an image signal processing pipeline that reduces noise, sharpens details, and adjusts colors to produce visually pleasing photos. While these processed photos are easy to collect and annotate, RAW images preserve richer light and intensity information and are often more useful for low-light vision, object detection, and autonomous driving. In this work, we rethink the task of recovering RAW images from ordinary photos. Instead of treating it as a simple pixel-to-pixel prediction problem, we view it as restoring the target RAW image distribution from a source distribution affected by complex perturbations introduced by the ISP pipeline. Based on this view, we study a single-step deterministic flow method, which learns the overall transformation from ordinary photos back to RAW images. Compared with standard regression methods, this approach is more robust to input perturbations and can better preserve real image structures. Building on this idea, we propose SHADE, which produces higher-quality RAW images and improves downstream visual tasks in real RAW scenarios.