FoMo: Forking Moment in Generative Trajectory as a Perceptual Distance
The paper uses diffusion “forking” time as an automatic perceptual-distance label for image pairs.
The authors argue that early splits in a diffusion trajectory mean images share only coarse structure, while late splits preserve most detail. They use that “forking moment,” FoMo, to train a reference-based image quality metric without human annotations. The paper says this produces pointwise labels that can compare arbitrary image pairs and outperforms human-annotated datasets across multiple benchmarks. HF Daily Papers' note
The authors argue that early splits in a diffusion trajectory mean images share only coarse structure, while late splits preserve most detail. They use that “forking moment,” FoMo, to train a reference-based image quality metric without human annotations. The paper says this produces pointwise labels that can compare arbitrary image pairs and outperforms human-annotated datasets across multiple benchmarks. HF Daily Papers' note
score 4