DanceOPD

DanceOPD (On-Policy Generative Field Distillation) is a training framework for flow-matching image generation models, from ByteDance and collaborators including the National University of Singapore's Tat-Seng Chua. It combines several capabilities, such as text-to-image generation, local editing, and global editing, in one student model. These capabilities often interfere with each other when trained jointly or merged.

Each frozen source model is treated as a velocity field. For each training sample, one field is chosen by hard routing. The field is queried at a single low-noise state taken from the student's own rollout, and the student is trained with a plain velocity MSE loss. Classifier-free guidance can be absorbed as another operator-defined field. The project page attributes the gains to three design choices: hard routing instead of soft teacher averaging, on-policy states, and a single query per rollout.

Reported results on the project page include 5.347 on GEditBench with 0.849 on GenEval for the T2I plus edit composition, and 5.498 / 0.848 for local plus global edit composition. These figures are the authors' own and are reported against baselines such as joint training, weight merging, off-policy distillation, DiffusionOPD and Flow-OPD. This is a research release. The project page does not link code, weights or a paper, and a YouTube review described the code as coming soon.