StableMotion: One-Step Motion Estimation with Diffusion Prior

A research team has introduced StableMotion, a framework that repurposes a pretrained text-to-image Stable Diffusion model as an image-to-motion estimator for single-image rectification tasks, delivering state-of-the-art results and a more than 100× speedup over prior diffusion-based methods [1][2]. The work, detailed in a paper by Ziyi Wang and colleagues, targets two specific computer-vision challenges: Stitched Image Rectangling (SIR) and Rolling Shutter Correction (RSC) [1][2]. The core innovation is treating a large-scale generative model not as an image creator but as a source of geometric and content priors for motion estimation. Diffusion models, which typically learn to reverse a process of adding Gaussian noise to data, have seen widespread commercial use in systems like Stable Diffusion and DALL-E [3]. StableMotion takes the Stable Diffusion backbone and repurposes it, moving from a text-to-image paradigm to an image-to-motion estimator [1][2]. A key finding enabling the framework's efficiency is what the authors term the Sampling Steps Disaster (SSD), a counterintuitive phenomenon where increasing the number of sampling steps in the diffusion process degrades output quality [1][2]. This observation directly motivated a one-step inference design. Supported by the SSD insight, StableMotion achieves its one-step inference, offering over a 100× speedup compared to previous diffusion model-based methods, even when combined with an optional Adaptive Ensemble Strategy (AES) post-processing step designed to consolidate multiple outputs into a single high-fidelity result [1][2]. The framework was evaluated on standard benchmarks and also showed promising transferability through qualitative examples and no-reference evaluations on unseen SIR-OOD and real-captured RSC benchmarks [1][2]. The paper was first submitted to the arXiv preprint server on May 10, 2025, with a revised version posted on June 26, 2026 [1]. The code and model weights have been made publicly available on GitHub [1][2].