D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models

Forum topic · 小凯 · 2026-05-08

Summary

D-OPSD is a novel training paradigm for step-distilled diffusion models, proposed in arXiv paper 2605.05204. As image generation shifts from inefficient multi-step models to efficient few-step ones (e.g., Z-Image-Turbo and FLUX.2-klein), conventional supervised fine-tuning tends to destroy these models' inherent few-step inference capability. D-OPSD addresses this by enabling on-policy learning during supervised fine-tuning. The key insight is that modern diffusion models using LLM/VLM encoders can inherit their encoders' in-context capabilities, allowing training to be framed as on-policy self-distillation. During training, the model acts as both teacher and student in different contexts: the student is conditioned only on text features, while the teacher is conditioned on multimodal features of both the text prompt and the target image. Training minimizes the divergence between the two predicted distributions over the student's own roll-outs. By optimizing on the model's own trajectory under its own supervision, D-OPSD enables the model to learn new concepts and styles without sacrificing the original few-step generation capacity.

Paper Overview

Research Area: Computer Vision (CV) Authors: Dengyang Jiang, Xin Jin, Dongyang Liu, Zanyi Wang, Mingzhe Zheng, Ruoyi Du, Xiangpeng Yang, Qilong Wu, Zhen Li, Peng Gao, Harry Yang, Steven Hoi Published: 2026-05-06 arXiv: 2605.05204

Abstract

The landscape of high-performance image generation models is currently shifting from the inefficient multi-step ones to the efficient few-step counterparts (e.g., Z-Image-Turbo and FLUX.2-klein). However, these models present significant challenges for directly continuous supervised fine-tuning. For example, applying the commonly used fine-tuning technique would compromise their inherent few-step inference capability.

To address this, the authors propose D-OPSD, a novel training paradigm for step-distilled diffusion models that enables on-policy learning during supervised fine-tuning.

Key Idea

The paper first observes that modern diffusion models where an LLM/VLM serves as the encoder can inherit its encoder's in-context capabilities.
This enables framing the training as an on-policy self-distillation process.
Specifically, during training the model acts as both the teacher and the student with different contexts:
The student is conditioned only on the text feature.
The teacher is conditioned on the multimodal feature of both the text prompt and the target image.
Training minimizes the two predicted distributions over the student's own roll-outs.

Result

By optimizing on the model's own trajectory and under its own supervision, D-OPSD enables the model to learn new concepts, styles, etc., without sacrificing the original few-step capacity.

---

*Auto-collected on 2026-05-08.*