On-Policy Self-Distillation for Multi-Turn Image Editing
Liangbing Zhao, Le Zhuo, Mohamed Elhoseiny
KAUST · Krea AI
Published September 28, 2026 · on the feed September 29, 2026 · ▲ 3 · arXiv 2609.35611
The 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.
The Brief
1:42A short podcast summary of the paper. Written and voiced by AI.
Transcript
On-Policy Self-Distillation for Multi-Turn Image Editing. By Liangbing Zhao and two others, from King Abdullah University of Science and Technology and Krea A I. Published on the twenty-eighth of September. Its code repository is public and real: about twenty-five source files.
Instruction-based image editing performs well when there is one instruction and one edit. Real editing is often iterative: each new instruction is applied to the output of the previous turn. Under that kind of recursive editing, the authors find that existing models degrade quickly. They attribute the failure to a mismatch between training and use. Models are trained on clean source images, but at inference they must keep conditioning on their own imperfect outputs, and those are not what they learned from.
MT-OPSD is an on-policy self-distillation framework built to close that gap. The model is trained on conditioning states it generated itself, the same kind of slightly degraded inputs it will meet in later turns.
Supervision for those states comes from a teacher conditioned on clean images, so the target stays sound even when the input is not. Because the states are produced by the model rather than collected, no multi-turn annotations are required. To measure long-horizon robustness, the authors also introduce LME-Bench, one hundred editing sessions of ten turns each.
Across three editing backbones, MT-OPSD substantially improves editing success over long horizons and reduces multi-turn collapse, while largely preserving single-turn editing quality.
A model that edits its own outputs needs to have practised on its own outputs. This framework gives it that practice, with a clean-conditioned teacher standing in for the ideal answer, and without asking anyone to annotate long editing sessions.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
One edit is easy. Ten in a row is where editors fall apart.
You use an AI editor to change a photo through several instructions: relight it, recolour it, move an object, restyle it. Each edit works on the previous result.
The road ahead
- 1 · The problem ~7 min Image editors trained on clean photos fall apart when each edit must start from their own slightly flawed previous output.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Let the editor practise on its own flawed outputs, with editing targets supplied by the same editor when it sees the clean photo.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Roll the model out to make flawed inputs, get targets from a clean-conditioned teacher, then train the student on those flawed inputs.
- 6 · The experiment ~6 min You’ll guess: Training an image editor on its own imperfect outputs reduced multi-turn collapse. What happened to its single-turn editing quality?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.