DanceOPD: On-Policy Field Distillation

• DanceOPD is an on-policy generative field distillation framework that models each image generation capability as a velocity field over a shared latent space.
• It employs hard routing with low-noise semantic queries to align the student’s rollout states with expert capability fields, addressing state-distribution mismatch.
• Evaluations on T2I, local/global editing, and CFG absorption demonstrate significant performance gains, highlighting the benefits of field-space routing and on-policy supervision.

Searching arXiv for the cited DanceOPD paper and closely related flow-matching distillation work. DanceOPD is an on-policy generative field distillation framework for flow-matching models that addresses multi-capability composition in image generation, including text-to-image (T2I), local editing, global editing, realism-field absorption, and classifier-free guidance (CFG) absorption (Zhou et al., 25 Jun 2026). It formulates each capability source as a velocity field over a shared flow state space, routes each sample to one capability field, queries one low-noise student-induced state, and trains the student with a simple velocity MSE objective (Zhou et al., 25 Jun 2026). Within this formulation, capability composition is not handled by parameter merging or naive joint supervision, but by learning from routed expert fields on the student’s own rollout states, thereby targeting target-field ambiguity, state-distribution mismatch, and trajectory-query correlation (Zhou et al., 25 Jun 2026).

1. Problem formulation and conceptual basis

Modern image generation systems are expected to support open-ended T2I generation, local editing, global editing, realism guidance, and CFG within a single deployed model, yet these capabilities are “not naturally aligned” and often conflict (Zhou et al., 25 Jun 2026). T2I favors open-ended diversity and prompt faithfulness, local editing emphasizes preservation of the source image with targeted changes, global editing seeks large transformations, realism fields push outputs toward photorealistic textures and lighting, and CFG modifies the generative field at inference time in ways that can harm diversity and stability when guidance is large (Zhou et al., 25 Jun 2026). A central problem is therefore capability composition: training one student flow model that strengthens target capabilities while preserving anchor generation quality (Zhou et al., 25 Jun 2026).

DanceOPD represents each frozen capability source as a velocity field on a shared latent state space,

$v_m(z_t, t, c),\quad m=1,\dots,M,$

where $z_t$ is the latent state at time $t$, $c$ is the conditioning, and $m$ indexes capabilities such as T2I, local edit, global edit, realism, or CFG (Zhou et al., 25 Jun 2026). The student $v_\theta(z_t,t,c)$ is another velocity field over the same space, and capability composition is cast as learning a single student field whose behavior matches the appropriate expert field on the states that the student itself will visit (Zhou et al., 25 Jun 2026). This shifts the problem from parameter-space compromise to field-space alignment.

A distinctive aspect of DanceOPD is its explicit rejection of off-policy supervision. Off-policy distillation queries teachers on fixed noised data or teacher trajectories, which creates a mismatch between training states and the states actually visited by the student at inference (Zhou et al., 25 Jun 2026). DanceOPD instead adopts an on-policy formulation: the student is rolled out from noise, the expert field is queried at a student-induced state, and the student is trained to match that routed expert at that location (Zhou et al., 25 Jun 2026). This suggests a view of generation in which the relevant supervision is defined by the student’s own trajectory geometry, not by an external noising distribution.

2. Flow-matching setting and routed on-policy objective

DanceOPD is built in the flow-matching setting, where generation is modeled as an ODE over a latent state $z_t$,

$\frac{d z_t}{dt} = v_\theta(z_t, t, c),$

integrated from a noise prior $z_T \sim p_T$ to $z_0$ (Zhou et al., 25 Jun 2026). Each capability source provides its own frozen velocity field $z_t$0 on the same latent space and scheduler (Zhou et al., 25 Jun 2026). For a routed sample, the supervising field is simply

$z_t$1

The DanceOPD loss is defined on a student-visited query state. Let the route be sampled as $z_t$2, the data be drawn from the corresponding capability distribution, the student rollout be $z_t$3, and the query time be selected via a semantic coordinate $z_t$4 mapped to $z_t$5 (Zhou et al., 25 Jun 2026). With $z_t$6, the objective is

$z_t$7

This is a simple velocity MSE objective, but its simplicity is inseparable from the routing and query policy (Zhou et al., 25 Jun 2026).

The paper also provides a local probabilistic interpretation. If student and teacher define local Gaussian reverse kernels,

$z_t$8

$z_t$9

then

$t$0

This implies that local KL-based velocity distillation is equivalent, up to a factor, to velocity MSE (Zhou et al., 25 Jun 2026). The choice of MSE is therefore not merely heuristic; it is compatible with a local kernel-matching interpretation.

3. Routing, query policy, and capability composition

DanceOPD imposes three coupled design choices: hard-routed sample-wise field matching, on-policy field querying, and semantic-side single low-noise query selection (Zhou et al., 25 Jun 2026). Hard routing means that each sample is routed to exactly one capability field rather than supervised by an average of multiple teachers (Zhou et al., 25 Jun 2026). The expectation over routes yields

$t$1

This preserves capability identity at the sample level (Zhou et al., 25 Jun 2026).

The rejected alternative is soft teacher mixing,

$t$2

which the paper characterizes as introducing target-field ambiguity because the resulting direction no longer corresponds to a meaningful single capability (Zhou et al., 25 Jun 2026). Empirically, hard routing improves GEditBench average by +15.2% versus soft mixing under MSE and by +10.6% versus soft mixing under a KL-weighted objective (Zhou et al., 25 Jun 2026). This is one of the clearest quantitative demonstrations in the paper that capability composition is more stable in field space when semantic roles are kept discrete rather than averaged.

The second design choice is query location. DanceOPD samples a semantic coordinate from a Beta distribution biased toward low-noise states,

$t$3

and queries only one such state per sample, with $t$4 (Zhou et al., 25 Jun 2026). The paper argues that low-noise states are semantically rich and therefore more informative for capability matching, while dense querying along a single trajectory produces highly correlated supervision (Zhou et al., 25 Jun 2026). Ablations show that single semantic-side query outperforms dense weighted variants $t$5 by up to 16.6%, while low-$t$6 queries improve GEditBench average over median-$t$7 by 23.7% and over high-$t$8 by 19.5% (Zhou et al., 25 Jun 2026). This suggests that query sparsity is not merely a computational convenience; it is structurally tied to reducing gradient correlation.

The third design choice concerns optimizer-step structure. The paper distinguishes step alternation, $t$9, from same-step accumulation, where gradients from multiple routed capability buckets are averaged in a single update (Zhou et al., 25 Jun 2026). Even when dense queries are absent, $c$0 degrades by about 4.6% relative to $c$1, and $c$2 degrades by 22.8% (Zhou et al., 25 Jun 2026). The interpretation offered is gradient-level capability conflict: even if routes are sampled separately, combining their gradients in one optimizer step partially recreates the interference that hard routing is meant to avoid (Zhou et al., 25 Jun 2026).

4. Algorithmic workflow and implementation details

A single DanceOPD training step consists of five operations (Zhou et al., 25 Jun 2026). First, one samples a route $c$3 and corresponding training pair $c$4. Second, one samples initial noise $c$5 and rolls out the student on-policy via

$c$6

Third, one samples a semantic coordinate $c$7, maps it to a time $c$8, and takes the stopped-gradient student state $c$9. Fourth, one queries the routed teacher $m$0. Fifth, one minimizes

$m$1

with respect to $m$2 (Zhou et al., 25 Jun 2026).

The main experiments use the Z-Image flow-matching model as backbone and train LoRA adapters on DiT blocks while freezing the base backbone parameters (Zhou et al., 25 Jun 2026). All capability sources share the same latent space, scheduler convention, and velocity parameterization, which the paper identifies as a prerequisite for shared field support (Zhou et al., 25 Jun 2026). Conditioning differs by capability: T2I uses text prompts, local editing uses source image, mask, and edit instruction, and global editing uses style or background-oriented prompts with image context (Zhou et al., 25 Jun 2026).

Training uses a 16-step Euler ODE rollout, AdamW, and learning rate $m$3 in Z-Image experiments (Zhou et al., 25 Jun 2026). Route probabilities are uniform over active capabilities. The paper gives explicit 1:1 routing between T2I and Edit for T2I-plus-edit composition, 1:1 between local-edit and global-edit for edit composition, and 1:1:1 among T2I, local, and global in diagnostics (Zhou et al., 25 Jun 2026). Rollout length is also ablated, with 16 steps reported as a good trade-off relative to 8, 20, and 28 steps (Zhou et al., 25 Jun 2026).

5. Capability fields, CFG absorption, and realism-field absorption

A central generalization in DanceOPD is that any operator-defined transformation of a velocity field can itself be treated as a capability field (Zhou et al., 25 Jun 2026). The most prominent example is CFG. With unconditional velocity $m$4 and conditional velocity $m$5, CFG at scale $m$6 defines

$m$7

DanceOPD can absorb this guided field by treating $m$8 as teacher and minimizing

$m$9

This is described as CFG absorption (Zhou et al., 25 Jun 2026).

The paper further analyzes how train-time absorption composes with inference-time CFG. If the student approximates the $v_\theta(z_t,t,c)$0-guided field during training and then external CFG with scale $v_\theta(z_t,t,c)$1 is applied at inference, the effective guidance is approximately multiplicative:

$v_\theta(z_t,t,c)$2

This means excessive train-time and test-time composition produces over-guidance (Zhou et al., 25 Jun 2026). Empirically, the best measured composition improves GEditBench average by +7.6% over train-only absorption and +1.4% over eval-only CFG, while excessive composition reduces performance by about 31.2% (Zhou et al., 25 Jun 2026). A plausible implication is that DanceOPD provides a principled way to internalize guidance, but only when guidance magnitude remains compatible across training and inference.

Realism-field absorption is treated analogously. A realism teacher specialized toward photorealistic textures and lighting is simply another frozen capability field $v_\theta(z_t,t,c)$3 (Zhou et al., 25 Jun 2026). On SD3.5-M, DanceOPD improves realism reward over off-policy distillation by 9.9%, closes 85.3% of the reward gap between base student and realism teacher, and keeps T2I score within 0.1% of off-policy distillation while remaining +7.6% above the student anchor (Zhou et al., 25 Jun 2026). The paper describes this as evidence that operator-like fields and reward-specialized fields can both be absorbed into the student without requiring separate inference-time modules.

6. Experimental results and ablation findings

The experimental evaluation is organized around T2I-plus-edit composition, local-plus-global edit composition, realism-field absorption, and CFG absorption (Zhou et al., 25 Jun 2026). For T2I plus Edit composition, DanceOPD improves GEditBench average over the best OPD baseline by 8.1%, improves over the Edit teacher itself by 8.5%, and slightly exceeds the T2I teacher on GenEval by 2.0% overall (Zhou et al., 25 Jun 2026). The reported category-level gains against DiffusionOPD include +21.9% on background change, +21.3% on style change, and +5.5% on color alteration (Zhou et al., 25 Jun 2026). These numbers support the claim that DanceOPD does not simply preserve an anchor capability; it can strengthen editing while maintaining or improving base generation quality.

For local plus global edit composition, the performance gap is larger. DanceOPD improves over the best competing composition baseline by 16.1% on GEditBench and over the local edit source by 7.9%, while its T2I score is 2.5% above the strongest composition baseline (Zhou et al., 25 Jun 2026). Category-level gains include +33.5% on background change, +12.9% on style change, and +11.6% on color alteration (Zhou et al., 25 Jun 2026). This is important because local and global editing are presented in the paper as especially conflicting capability fields (Zhou et al., 25 Jun 2026). The result suggests that field-space routing and on-policy querying are particularly beneficial when capabilities are semantically orthogonal.

The ablations isolate the mechanism of the framework. Plain MSE outperforms timestep-weighted MSE, DMD-EMA, SDS+DMD, KL-weighted, consistency, DMD2, and AuxFeat by 2.8–4.5% in balanced performance (Zhou et al., 25 Jun 2026). Hard routing outperforms soft mixing; low-noise semantic-side queries outperform median- and high-noise queries; and $v_\theta(z_t,t,c)$4 outperforms dense-query and same-step-accumulation settings (Zhou et al., 25 Jun 2026). SDE decorrelation partially rescues dense-query degradation, with an 18.4% improvement in stress settings, but still underperforms the default single-query configuration by about 8.6% (Zhou et al., 25 Jun 2026). These findings collectively indicate that the performance gains come less from architectural novelty in the backbone and more from the precise structure of routed, on-policy, low-noise field supervision.

Initialization is another strong factor. For edit composition, initializing from the local-edit teacher substantially outperforms merged, global-edit, or T2I-only initializations, with improvements of up to +112.8% over global-edit init on GEditBench average and +204.4% over T2I init in some settings (Zhou et al., 25 Jun 2026). The paper interprets this as anchor preservation through initialization plus route balancing rather than through explicit regularization terms (Zhou et al., 25 Jun 2026). This suggests that DanceOPD is not a plug-and-play multi-task recipe independent of starting point; its best performance depends on choosing an initialization that already reflects the desired anchor capability.

7. Relation to prior work, limitations, and interpretation

DanceOPD is situated within on-policy distillation for flows and diffusion, but is distinguished by its emphasis on multi-capability composition rather than sampler compression or single-task transfer (Zhou et al., 25 Jun 2026). The paper contrasts it with DiffusionOPD, which uses dense supervision on all timesteps and a transition-mean or KL-style objective, and with Flow-OPD, which uses on-policy SDE rollout and PPO-style RL (Zhou et al., 25 Jun 2026). DanceOPD instead uses hard routing, single low-noise query, and simple MSE without PPO (Zhou et al., 25 Jun 2026). The broader claim is that properly chosen query semantics can substitute for more elaborate optimization machinery.

Several limitations are explicit. All capability sources must share the same latent space, scheduler, and velocity parameterization, which restricts application across heterogeneous teacher families (Zhou et al., 25 Jun 2026). Routing is predefined rather than learned, so ambiguous or mixed-capability prompts would require an additional routing mechanism (Zhou et al., 25 Jun 2026). On-policy rollout increases training cost relative to off-policy methods, even though DanceOPD reduces cost by using only one gradient-bearing query per rollout (Zhou et al., 25 Jun 2026). The student also remains bounded by teacher quality; a poor or misaligned capability field constrains what can be absorbed (Zhou et al., 25 Jun 2026). Finally, the paper studies image generation only, though it states that extension to other modalities is conceptually straightforward (Zhou et al., 25 Jun 2026).

A plausible implication is that DanceOPD’s main contribution is less a new generative architecture than a training theory for composing expert fields in a shared flow space. Its emphasis on routed semantics, on-policy state support, and sparse low-noise querying amounts to a specific answer to the question of how one should supervise a student when the target is not a single teacher, but a family of partially conflicting operators and models. In that sense, DanceOPD is best understood as a generative field distillation framework for capability unification rather than as a domain-specific image generation model.