On-Policy Self-Distillation for Multi-Turn Image Editing

2026-09-28

AI Summary

This study addresses the performance degradation in multi-turn image editing caused by distribution shifts between training and inference, proposing the MT-OPSD framework. This method introduces a novel unsupervised online policy self-distillation mechanism that leverages a clean-condition teacher model to provide supervisory signals, enabling iterative training on self-generated states and enhancing long-horizon robustness without requiring multi-turn annotated data. Furthermore, LME-Bench, a benchmark comprising 100 ten-turn conversational sessions, is constructed to systematically evaluate long-term editing capabilities. Experimental results demonstrate that the proposed framework significantly improves long-horizon editing success rates and effectively mitigates multi-turn collapse while preserving single-turn editing quality.

Abstract

Instruction-based image editing has achieved strong performance in single-turn settings, yet practical editing is often iterative, with each instruction applied to the output of the previous turn. We find that existing editing models degrade rapidly under recursive editing and attribute this failure to a train-test mismatch in the conditioning distribution: models are trained on clean source images but must repeatedly condition on their own imperfect outputs at inference time. To address this, we propose MT-OPSD, an on-policy self-distillation framework that trains the model on self-generated conditioning states with editing supervision from a clean-conditioned teacher, without requiring multi-turn annotations. We further introduce LME-Bench, a benchmark of 100 ten-turn editing sessions for evaluating long-horizon robustness. Experiments across three editing backbones show that MT-OPSD substantially improves long-horizon editing success and reduces multi-turn collapse while largely preserving single-turn editing quality.

Problem

• multi-turn image editing
• instruction-based image editing
• train-test mismatch
• recursive editing degradation

Innovation

• On-Policy Self-Distillation
• Multi-Turn Image Editing
• Train-Test Mismatch
• Long-Horizon Robustness
• LME-Bench