Key Features
Composes multiple image-generation capabilities through on-policy field distillation.
Treats each source capability as a frozen velocity field.
Uses hard-routed sample-wise field selection instead of soft averaging teachers.
Queries teacher fields on student-visited rollout states.
Uses one low-noise semantic query to reduce correlated supervision.
Trains with simple local velocity MSE matching.
Can absorb operator-defined fields such as classifier-free guidance.
Provides arXiv manuscript and public code link.
About DanceOPD
The method addresses target-field ambiguity, state-distribution mismatch, and trajectory-query correlation. It uses hard routing to select one frozen capability field per sample, a single low-noise semantic-side query on a stop-gradient student rollout state, and local velocity MSE matching.
DanceOPD is useful for unifying text-to-image generation, local editing, global editing, style absorption, and classifier-free-guidance absorption without degrading anchor capabilities. The project provides a manuscript, code link, diagnostics, ablations, and qualitative galleries.