FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance

The paper proposes a fully automated pipeline that generates pointwise perceptual distance labels between image pairs based on the generative dynamics of diffusion models without human annotation. The core idea exploits how coarse image structures are generated in early timesteps and fine details in later timesteps. The generation forking moment, named FoMo, is used as a reference-grounded distance label to supervise training of an image quality assessment metric. Experiments confirm this approach outperforms human-annotated datasets across multiple benchmarks.