On-Policy Self-Distillation for Multi-Turn Image Editing
MT-OPSD uses on-policy self-distillation so image editors stay stable across iterative multi-turn edits.
MT-OPSD is an on-policy self-distillation method for instruction-based image editing that trains models on their own imperfect intermediate images while a clean-conditioned teacher provides editing supervision, avoiding multi-turn annotations. The authors introduce LME-Bench, with 100 ten-turn editing sessions, to measure long-horizon robustness. Across three editing backbones, the method raises multi-turn success and reduces recursive collapse while largely preserving single-turn quality.
- Multi-turn failure comes from training on clean images but conditioning on imperfect outputs.
- On-policy self-distillation uses a clean-conditioned teacher without multi-turn labels.
- LME-Bench contains 100 sessions of ten sequential editing turns.
- Improves long-horizon success on three backbones while keeping single-turn quality.
From the original article
1 min read at huggingface.co
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.
By Liangbing Zhao, Le Zhuo, Mohamed Elhoseiny · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.