AIAny
Icon for item

Self-OPD: On-Policy Distillation for Flow Matching Models without Teacher

Turns a flow-matching image generator's self-exploration into dense, per-step supervision without a pretrained teacher; it branches the student's next-state into stochastic SDE candidates, scores them against a deterministic self-reference, and applies an advantage-weighted pull–push velocity regression with reward-level fusion for multi-objective alignment.

Introduction

Training task-specific teacher models for on-policy distillation incurs high compute and can propagate teacher biases into student trajectories. Self-OPD sidesteps that dependency by converting the student’s own local stochastic exploration into dense, stepwise supervision: at each denoising step it samples K SDE branches from the student, rolls them to terminals, scores them versus a deterministic ODE baseline, and uses normalized advantages to shape the velocity field.

Key Findings
  • Teacher-free per-step supervision: Replaces teacher velocity targets with advantage-weighted velocity targets discovered via the student's local branches, avoiding the cost and bias of training specialized teachers.
  • All-branch pull–push objective: High-advantage branches attract the velocity field while low-advantage branches repel it; direction-aware attenuation and SDE-variance normalization stabilize learning and connect the loss to a reward-tilted KL interpretation.
  • Reward-level fusion for multi-objectives: Normalizes and fuses scalar rewards before ranking branches, avoiding direct gradient conflicts across objectives and enabling a single image to satisfy multiple metrics simultaneously.
  • Empirical gains: On single- and mixed-reward benchmarks, Self-OPD matches or outperforms prior RL and teacher-based OPD approaches while remaining teacher-free.
Who It's For and Tradeoffs

Great fit if you work on aligning flow-matching or continuous denoising models to black-box, possibly heterogeneous rewards and want dense, stable per-step supervision without training task-specific teacher networks. Look elsewhere if your priority is minimal sampling overhead at each step—Self-OPD requires K stochastic branches and deterministic rollouts per timestep, which raises per-step compute and evaluation costs compared to purely terminal RL approaches.

Where It Fits

Positions itself between teacher-based OPD (which provides dense supervision but needs pretrained experts) and terminal-policy RL (which is sample-inefficient and sparse): Self-OPD keeps dense, on-policy feedback while removing external teachers, trading extra local sampling for reduced dependence on specialized teacher models and better multi-objective joint optimization.

Information

  • Websitearxiv.org
  • OrganizationsTsinghua University, Zhejiang University, Alibaba Group
  • AuthorsShiyi Zhang, Mushui Liu, Yunze Tong, Wanggui He, Siyu Zou, Jinlong Liu, Yunlong Yu, Jian Song, Hao Jiang, Pipei Huang
  • Published date2026/08/27

More Items

Provides a real-scale 3D Hong Kong sandbox to evaluate whether multimodal LLM agents can turn local street-view perception into sustained spatial action, supporting closed-loop first-person interaction, an interactive map, and controlled tests of grounding, long-range navigation, and robustness.

Measures whether video generators reproduce the correct distribution of possible physical behaviors under repeated rollouts. Introduces PAWBench and PAWEval to convert repeated generations into outcome-level empirical distributions and quantify probabilistic alignment; evaluates 50 scenarios and 11 models and finds no model consistently matches reference probabilities.

Converts image-level rewards into explicit intermediate targets for diffusion-model denoising via an on-policy self-distillation loop. Constructs bounded positive/negative targets around anchors from reward gradients, fits those targets with finite updates, and refreshes a behavior policy by EMA—improving aligned performance across backbones while reducing GPU hours.