When Preference Labels Fall Short: Aligning Diffusion Models from Real Data

Researchers studied whether real images can supervise preference alignment for diffusion models.

Instead of relying on labeled preference pairs from generated samples, they contrast real references with generated or perturbed outputs.

The data-centric approach works effectively and can match standard preference-based alignment methods.

This points to a simpler, more label-efficient way to steer generative models using existing real data.