Any-OPD proposes heterogeneous on-policy distillation between latent flow-matching generators whose VAEs, architectures, and timestep schedules differ. It compares independently decoded teacher and student outputs in a frozen, model-agnostic vision representation, aligns trajectories by continuous noise levels, and adds a short anchoring phase using teacher samples re-encoded by the student VAE. The supplied abstract reports that distilling 12B FLUX.1-dev into 2.5B SD3.5-Medium raises PickScore from 0.846 to 0.884 and HPSv3 from 9.12 to 10.97. However, the provided publication date is future-dated, so the paper and claims require verification.
No heat snapshots are available in the last 24 hours.