DanceOPD trains one image model to do text-to-image, local editing, and global editing without those capabilities canceling each other out. Each training sample is hard-routed to exactly one frozen teacher Velocity field, queried on the student’s own rollout at a single low-noise step, and matched with plain MSE. Hard routing beats averaging multiple teachers by 15.2% on edit benchmarks.
You’ve shipped an image generation product. Users want one model that can generate from a prompt, do precise local edits (“change the apple to crystal”), and do global restyling (“make it cyberpunk”). You have a great T2I checkpoint and two specialist edit checkpoints. Today the options are bad: train jointly and watch capabilities dilute each other, merge weights and get a compromise model worse than any input, or stack them at inference time and pay 3x. The dominant research baseline here is Off-policy distillation, where you noise the target image and ask teachers to supervise that noised state. The problem: the deployed student visits different states than those noised endpoints, so training and inference diverge.
The paper’s framing: each frozen teacher (T2I, local-edit, global-edit, even classifier-free guidance) is a Velocity field over the same shared state space used by Flow matching models. Distilling them into one student becomes three coupled questions. Which teacher supervises this sample? Where do we query that teacher? How many states from the student’s trajectory get supervised?
DanceOPD answers each one in the simplest possible way. First, hard route: each sample is dispatched to exactly one teacher based on its data partition. No averaging, no soft mixing. Second, query On-policy distillation: roll out the current student for 16 steps, pick a state from that rollout (stop-gradient so you don’t backprop through the solver), and ask the routed teacher what velocity it would predict at that state. Third, sample one state per rollout, biased toward the low-noise end of the trajectory where capability-specific signal is densest. Loss is plain velocity MSE between student and teacher prediction. The paper shows in an appendix that KL between locally-Gaussian transition kernels reduces to weighted MSE, so MSE is the natural objective here, not a shortcut.
for step in training:
m = sample_route(active_buckets) # one teacher
x, c = sample_from(D[m])
z_T = sample_noise()
traj = student.rollout(z_T, c, steps=16) # stop-grad
s = beta_sample(5, 2) # low-noise bias
z_t = stop_grad(traj[t_of(s)])
target = teachers[m](z_t, t, c) # frozen
loss = mse(student(z_t, t, c), target)
loss.backward()
The same recipe absorbs Classifier-free guidance (CFG) by treating the guided velocity as just another teacher field.
The prevailing instinct when distilling multiple capabilities is to average teacher signals inside one update (soft mixing) or to harvest many supervision points from one rollout (dense queries) so you get more signal per expensive student rollout. This paper shows the opposite: averaging teachers destroys the semantic identity of each capability, and dense queries from one rollout are so correlated they double-count rather than diversify. The cleanest evidence is the diagnostic comparison where hard routing beats soft all-teacher mixing by 15.2% under MSE and 10.6% under KL, isolating routing from objective.
The load-bearing finding is the routing and query diagnostics, because they prove the mechanism rather than just a leaderboard lift. Hard routing beats soft all-teacher mixing by 15.2% (MSE) and 10.6% (KL) on GEditBench-EN. Querying at low-noise semantic-side states beats median-noise and high-noise queries by 23.7% and 19.5%. Adding more queries per rollout doesn’t help: K=2,4,8,16 dense variants all underperform the single-query default by 7.9% to 16.6%. When forced to use dense correlated queries, switching from ODE to SDE rollout recovers 18.4% of the loss, confirming trajectory correlation is the actual culprit, not on-policy supervision itself.
On the headline composition tasks, DanceOPD beats the strongest competing on-policy baseline (DiffusionOPD) by 8.1% on edit benchmarks while slightly improving T2I quality on GenEval. For local + global edit composition the gap widens to 16.1% over the best competing baseline. Realism-field absorption closes 85.3% of the student-to-teacher reward gap while keeping T2I within 0.1% of off-policy distillation. A practical caveat the paper surfaces: absorbed and inference-time guidance compose multiplicatively (effective scale ≈ α·β), so naively combining them over-guides and tanks scores by 31.2%.
Reach for this when you have several specialist generators (a T2I checkpoint, an edit checkpoint, a style checkpoint) and want one deployed model that does all of them without the capability-interference tax. The recipe is concrete: keep your data partitions labeled by capability, route each batch sample to exactly one teacher, generate the query state by rolling out the current student (not by noising the target image), pick one low-noise step per rollout, and regress velocity with MSE. Initialize from the strongest capability-relevant checkpoint, not from a merged anchor (the paper shows local-edit init beats merged init by 37.2% at 2k steps).
The authors maintain a project page. The paper experiments on Z-Image and SD3.5-M as backbones with LoRA adapters at rank 128. The realism reward model used for evaluation is proprietary and not released. The paper does not mention a code release at submission time, and the DiffusionOPD baseline had to be reimplemented because that code wasn’t public either.
When composing capabilities, preserve each sample’s semantic identity and supervise where the student actually goes. Averaging teachers feels like the principled choice but produces a target direction that corresponds to no real capability. Querying many points along one rollout feels efficient but the points are correlated, so you pay compute without buying independent signal. One teacher, one query, the student’s own state.
•
The frozen teachers must share a backbone family, latent space, scheduler, and velocity parameterization. The paper is explicit about this. If your candidate teachers come from different families, none of this applies without nontrivial bridging.
•
Hard routing assumes you can label each training sample with one capability. When a single prompt legitimately needs two capabilities at once (e.g., “edit the background AND restyle”), the routing assumption weakens and the paper offers no learned router, only data-partition routing.
•
Several of the most striking ablation numbers (e.g., the 15.2% hard-vs-soft gap, the 23.7% low-vs-median timestep gap) come from controlled diagnostics on GEditBench-EN averages, not from the main composition tables. They isolate the mechanism cleanly but should not be read as expected lifts on arbitrary edit workloads.