Timing Matters Twice: When to Query and When to Skip in Flow-Matching On-Policy Distillation

Renye Yan, Junyi Zhan, Jikang Cheng, Wei PengarXiv 2026 · arXiv

TL;DR

PhaseOPD is a new framework for on-policy distillation that adaptively determines the optimal moment to query specialized teachers and terminates student rollouts early once feedback is obtained. By aligning teacher supervision with the natural phases of the generative process, it slashes training time by over 40% while simultaneously boosting image quality, semantic structure, and text rendering accuracy.

What Problem Does This Paper Address?

Post-training generative models often involves On-Policy Distillation (OPD), where a student model learns from specialized "teacher" models (experts in aesthetics, OCR, etc.) by observing its own generated samples. However, current methods typically use a one-size-fits-all schedule for these queries.

This is inefficient for two reasons. First, different capabilities emerge at different times: global structure forms early, while fine textures appear late. Querying a "detail" teacher too early can actually disrupt the formation of global structure. Second, generating a full 16-step or 50-step image trajectory just to get feedback at step 5 is a massive waste of compute. The authors argue that "timing matters twice"—we need to know exactly when to query and when to stop calculating.

Methods

The authors introduce PhaseOPD, which identifies three distinct generative phases online: Coarse Layout, Content Formation, and Fine-scale Revision.

  1. Online Phase Perception: Instead of extra model evaluations, PhaseOPD analyzes the student's existing velocity predictions. It calculates a "spatial-scale revision ratio" ($F_k$) to measure how much the estimated clean image changes between steps at different scales.
  2. Adaptive Boundary Detection: By tracking the rate of change in these revisions ($\Delta F_k$), the system identifies "phase boundaries" where the generative process shifts gears (e.g., from layout to detail).
  3. Teacher-Conditioned Routing: Specialized teachers are queried only during their "preferred" phase. For instance, a structure-focused teacher (GenEval) is queried in Stage I, while an aesthetics teacher is queried in Stage III.
  4. Dynamic Truncation: This is the "skip" part. Once the necessary teacher queries for a batch are finished, the remaining denoising steps are cancelled, saving significant GPU time.
Figure 2. Overall framework of our method

Datasets

The models were trained using prompts from Pick-a-Pic (aesthetics) and splits from FlowGRPO (GenEval and OCR). Evaluation relied on standard benchmarks like DrawBench, GenEval, and OCR accuracy. To ensure the model wasn't just "reward hacking" its teachers, the authors also used held-out diagnostic metrics like ImageReward and VQAScore on the GenAI-Bench-1600 dataset.

Key Insights and Ideas

  • Phase-Specific Utility: The paper provides empirical evidence that teacher feedback has a "sweet spot." Structure-related supervision is most effective in early steps, whereas content and appearance benefit more from middle and late steps.
  • Interference Reduction: By not forcing every teacher to supervise every step, PhaseOPD reduces "cross-stage interference." A detail-oriented teacher won't accidentally "over-constrain" the model while it's still trying to figure out the basic layout.
  • Compute is the Bottleneck: In multi-teacher OPD, the rollout (generating the image) is often the most expensive part. PhaseOPD's ability to skip the "rollout suffix" is a practical breakthrough for scaling these methods.

Key Results

  • Efficiency: On SD3.5-M, PhaseOPD only needed to execute ~5.6 out of 16 rollout steps on average. This led to a 41.3% reduction in training time compared to the state-of-the-art DanceOPD.
  • Quality: Despite training faster, it achieved higher scores across the board. As shown in the training curves, it reaches high quality levels much earlier in the training process than baselines.
  • Generalization: The method proved "plug-and-play," improving performance when added to different base models like FLUX.1-dev and BAGEL.
  • Human Preference: In double-blind studies, users preferred PhaseOPD's outputs over those from standard OPD methods.
Figure 3. Quality over the first 160 training minutes. Pan- els show Aesthetic, GenEval, and OCR for DanceOPD, Diffu- sionOPD, and Ours. Time uses each method’s matched mean up- date time measured with exclusive GPU access.
Figure 5. Qualitative comparison with the primary teacher set. Rows correspond to Aesthetic, GenEval, and OCR; columns show SD3.5-M Base, DanceOPD, DiffusionOPD, and Ours. All meth- ods use archived, paired outputs generated with the same seed.

Limitations and Future Work

While PhaseOPD significantly improves text rendering, it isn't perfect; complex punctuation or very dense text can still fail. Additionally, while the timing of queries is adaptive, the assignment of teachers to specific phases (e.g., "OCR goes in Stage II") is currently a manual design choice based on task intuition. Future work could explore ways to automatically discover these phase preferences for new, unknown teacher models.

Summary updated

Abstract

On-policy distillation (OPD) enables flow-matching models to acquire complementary capabilities from specialized teachers through supervision on student-generated trajectories. However, different evaluation dimensions, such as aesthetics and OCR accuracy, and different tasks, such as image generation and editing, emphasize different aspects of the denoising process, suggesting that their effective supervision may occur at different stages. Existing methods typically select supervision states from predefined timestep distributions, which may overlook these phase-dependent differences and incur unnecessary computational overhead. This motivates two critical timing decisions: when to adaptively query teachers for effective supervision based on the evolving generative process, and when to skip the remaining rollout to avoid redundant computation. In this paper, we introduce PhaseOPD, a phase-aware query-and-skip framework that identifies generative phases online from revisions of predicted-clean latents, using existing student velocity predictions without additional model evaluations. For each task-assigned teacher, PhaseOPD selects a supervision state based on its phase preference and skips the unused rollout suffix once all query decisions are finalized, preserving the single-query velocity-distillation objective. On SD3.5-M, PhaseOPD executes only 5.56 of 16 rollout positions per local batch, reducing training time by 41.3% relative to DanceOPD while improving Aesthetic, GenEval, and OCR scores.
BibTeX
@misc{yan2026timing,
  author = {Renye Yan and Junyi Zhan and Jikang Cheng and Wei Peng},
  title = {{Timing Matters Twice: When to Query and When to Skip in Flow-Matching On-Policy Distillation}},
  year = {2026},
  eprint = {2610.34191},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.34191},
}
Figure