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

Research on iterative image editing has moved from single-turn success toward a harder practical problem: how a model behaves after many consecutive edits. MT-OPSD addresses this by training on the model's own edited states, while LME-Bench tests long-horizon robustness with 100 ten-turn editing sessions. For users who refine images through repeated prompts, the result matters because a system can look reliable after one command and still lose control over the image after a short chain of follow-ups.

From Single-Turn Wins to Multi-Turn Failures

Instruction-based image editing tools are usually evaluated after a single prompt, such as changing background, lighting, or object placement. That setting is useful, but it does not match many real workflows. In practice, people often ask for one change, inspect the output, then request another change based on the first result. This creates a recursive editing process, where each instruction is applied to the previous output rather than to the original image.

The source material describes a clear failure pattern: models that perform well in single-turn editing can degrade rapidly under recursive editing. The breakdown is not just a small quality drop. It suggests that the conditioning chain itself becomes unstable once the model must repeatedly take its own imperfect results as input.

Why Recursive Editing Breaks the Training Setup

The explanation given for this failure is a train-test mismatch in the conditioning distribution. During training, many image editing models are shown clean source images and then asked to produce an edited target. At inference, however, a multi-turn session forces the model to condition on its own generated images, which may contain artifacts, misplaced details, or other errors that were absent during training.

This is a familiar kind of distribution shift in sequential generation tasks. The model learns from a stable input distribution but operates in a changing one. In iterative image editing, the problem is especially visible because visual errors can accumulate. A small artifact in turn one may be reinforced in turn two, making later instructions harder to follow.

MT-OPSD and LME-Bench Move the Test to Ten Turns

MT-OPSD responds to this mismatch with on-policy self-distillation. In this framework, the model is trained on self-generated conditioning states, meaning the inputs it learns to handle are closer to the outputs it will actually see in a multi-turn session. Editing supervision comes from a clean-conditioned teacher, and the method does not require multi-turn annotations.

The evaluation also introduces LME-Bench, a benchmark of 100 ten-turn editing sessions designed to measure long-horizon robustness. This shifts the test from "can the model follow one instruction?" to "can the model remain reliable through a sequence of edits?" Experiments across three editing backbones report that MT-OPSD substantially improves long-horizon editing success and reduces multi-turn collapse while largely preserving single-turn editing quality. The project page is https://liangbingzhao.github.io/MT-OPSD/.

What the Latest Results Suggest

The current stage of this work is diagnostic as well as constructive. It identifies a specific failure mode in iterative image editing and proposes both a training method and a benchmark for testing it. The key numbers are concrete: 100 ten-turn editing sessions in LME-Bench, three editing backbones in the reported experiments, and a method trained on self-generated states without multi-turn annotations.

The next signal will come from whether long-horizon editing tools can move beyond benchmark evaluation and maintain stability in everyday creative workflows. The research points to a clearer requirement: robustness should be measured across sequences, not only after the first prompt.

Meta Description:
MT-OPSD and LME-Bench tackle ten-turn iterative image editing collapse by training on self-generated states with teacher supervision across three backbones.

Tags:
iterative image editing, MT-OPSD, LME-Bench, on-policy self-distillation, long-horizon robustness

Featured image:
abstract technology style, no real people, no company logos, no text