Reference-based image quality assessment (IQA) aims to quantify how humans perceive the difference between two images. Training such metrics typically depends on human-annotated data, but the two dominant labeling schemes each carry drawbacks: mean opinion score (MOS) pointwise scoring is convenient to annotate yet expensive to collect at scale and noisy due to inconsistent judgments, while two-alternative forced choice (2AFC) pairwise labels are reliable and efficient but capture only relative comparisons.
To sidestep human annotation entirely, researchers propose FoMo (Forking Moment), a fully automated data generation pipeline that produces pointwise perceptual distance labels for image pairs. The method treats the generative dynamics of diffusion models as a proxy for perceptual distance, exploiting the observation that coarse image structure emerges in early timesteps while fine details appear later.
Under this view, two images that diverge — or "fork" — early in the generation process share only coarse structure and are therefore perceptually far apart, whereas images that fork late differ only in fine detail. The authors report that this diffusion trajectory aligns well with the human visual system, and use the forking moment as a reference-grounded distance label to supervise a reference-based IQA metric.
Because the resulting labels are pointwise, they support universal comparison between arbitrary image pairs, enabling what the authors describe as an information-rich training objective. Extensive experiments across diverse backbone architectures reportedly confirm the effectiveness of the generation pipeline, outperforming human-annotated datasets on multiple benchmarks.