Image Rotation Correction via Diffusion and Consistency Models

Tilted images are common in everyday photography and pose challenges for traditional correction methods that rely on geometric transformations or prior knowledge of rotation angles. Existing approaches, such as mesh warping or optical flow prediction, either require accurate structural cues or suffer from instability in complex scenes. Recent advances in generative modeling, particularly diffusion models, offer strong priors but remain computationally intensive. In this paper, we propose a novel one-step image rotation correction framework based on Consistency Models, a new class of generative models that enable efficient, angle-free, and semantically consistent transformation. Our method learns a direct mapping from a tilted image to its upright version using self-supervised consistency constraints, bypassing the need for explicit angle estimation or flow computation. Experimental results demonstrate that our approach achieves superior visual quality and inference speed compared to state-of-the-art methods, especially in large-tilt and structure-free settings.