On-Policy Self-Distillation for Multi-Turn Image Editing
The paper targets a specific failure mode: image editors falling apart when they must edit their own prior outputs.
The authors say current instruction-based editors are trained on clean source images but face degraded, self-generated images during iterative use. MT-OPSD trains on those self-generated conditioning states, using a clean-conditioned teacher for supervision and no multi-turn annotations. They also introduce LME-Bench, with 100 ten-turn editing sessions, to test long-horizon robustness. Across three editing backbones, the method improved multi-turn success while mostly preserving single-turn quality. HF Daily Papers' note
The authors say current instruction-based editors are trained on clean source images but face degraded, self-generated images during iterative use. MT-OPSD trains on those self-generated conditioning states, using a clean-conditioned teacher for supervision and no multi-turn annotations. They also introduce LME-Bench, with 100 ten-turn editing sessions, to test long-horizon robustness. Across three editing backbones, the method improved multi-turn success while mostly preserving single-turn quality. HF Daily Papers' note
score 4