MT-OPSD lifts 10-turn success from 3–15% to 38–52% without multi-turn labels

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

Liangbing Zhao, Le Zhuo, Mohamed Elhoseiny

cs.CV, cs.LG

2026-09-29

MT-OPSD uses identity rollouts and on-policy self-distillation to lift 10-turn success from 3–15% to 38–52% on three editors, keeping collapse ≤4% and ImgEdit intact.

What problem this solves

Instruction-based image editors already work for a single turn: lighting, style, object moves. Real workflows chain those edits, each turn feeding on the last output. Off-the-shelf models fall apart under that recursion. A few turns later the picture picks up chromatic speckle, broken structure, or identity drift.

KAUST and Krea AI pin this on a train-test mismatch in the conditioning distribution, the distribution over images the model is asked to edit. Training always conditions on clean sources. At test time the model has to condition on its own slightly wrong outputs, and those errors compound. A blunt diagnostic makes the point: tell the model to leave the image unchanged, and Qwen-Image-Edit and OmniGen2 still drift turn after turn. Same family of failure as exposure bias in autoregressive video generation, where training sees clean frames and inference steps on the model's own predictions.

Multi-turn labeled sequences are scarce. Backpropagating through hundreds of denoising steps stacked across turns is too expensive. Training-free patches such as Emu Edit copy nearly unchanged pixels back to the previous input, which helps local edits and fails when a restyle or relight is supposed to change most of the frame.

Method

MT-OPSD starts from a concrete claim: a pretrained editor already knows how to edit under clean conditioning. The missing piece is to keep editing once the condition itself is a self-generated, slightly corrupted frame. The method never models full multi-turn histories. It isolates model-induced error with identity rollouts.

Four moving parts:

• Identity rollout. From a clean source, repeatedly apply Make everything unchanged with the same sampler used at inference. Content stays put; the residual is mostly the model's own error. Rolling out with real edit instructions would mix intended semantic change into that residual.
• Two-branch objective, sampled 2:1 edit vs identity. The identity branch treats the rollout state as its own target and only asks the model not to drift further. Using the clean source as a restoration target, fixing k turns of error in one step, proved too hard to optimize. The editing branch pairs the same dirty state with a real instruction. The student conditions on the dirty image, a frozen teacher conditions on the clean source, and sparse velocity matching aligns them at noisy latents the student actually visited. That is on-policy self-distillation: the student is supervised on its own states, the teacher under a richer condition.
• Adaptive curriculum. Rollout depth increases by one only when mean pixel drift stays under a threshold for a patience window, capped at Kmax=10. In practice the curriculum saturates around four turns.
• Gated teacher promotion. A checkpoint is dumped every 50 samples. GPT-4o scores 23 held-out ten-turn sessions; the teacher is replaced only when long-horizon success and collapse improve. The VLM is a gate, not a training reward.

Training uses about 2,000 OmniEdit source-instruction pairs. Ground-truth edited images and multi-turn annotations are unused. Three backbones get LoRA rank 32: Qwen-Image-Edit-2511, FireRed-Image-Edit, and FLUX.2-klein-base. Four H100/H200 GPUs train, four more run the async gate.

Results

LME-Bench holds 100 sessions of 10 turns. Sources are 1024x1024 Z-Image-Turbo images, split across 10 semantic categories. Each session has six local and four global edits, with globals non-adjacent and kept out of the last two turns. GPT-4o scores prompt following and consistency from 0 to 10; a turn counts as success only if both are at least 7. SR@k is the fraction of sessions with every one of the first k turns successful. CR@k is the fraction that have already suffered two consecutive visually degraded turns by turn k.

Open-source bases drop to SR@10 of 0.03-0.15 and CR@10 of 0.25-0.61. Training-free helpers are backbone-specific: Emu Edit helps Qwen most, VAE-LFA helps FireRed more, and FreqEdit usually hurts at long horizon. MT-OPSD lifts SR@10 to 0.38-0.52 and holds CR@10 to at most 0.04, without giving up early-turn success. The curriculum only reached about four turns; the ten-turn gains still show up.

Among closed models, GPT-Image-1 almost fully collapses by turn 10 (SR@10=0.01, CR@10=0.69) and loses to MT-OPSD on every reported turn. Nano Banana rarely collapses; its SR is dragged down by early global edits (SR@10=0.56). GPT-Image-2 and Nano Banana Pro remain ahead at long horizon (0.96 and 0.91).

On MSE-Bench, a five-turn suite dominated by local edits, Qwen SR@5 goes from 0.24 to 0.51, FireRed from 0.40 to 0.69, and FLUX stays at 0.49. Single-turn ImgEdit barely moves: Qwen 4.51 to 4.49, FireRed 4.56 to 4.52, FLUX 4.20 to 4.28.

Ablations on Qwen lock the mechanism in. Drop the editing branch and SR@10 becomes 0.00, CR@10 0.80. Drop the identity branch and SR@10 is still 0.49, but the model over-edits and ImgEdit falls from 4.49 to 3.88. Replace on-policy velocity matching with ordinary flow matching onto teacher-generated images and SR@10 returns to 0.03, CR@10 0.67, worse than the base. A fixed four-turn depth yields SR@10 of 0.17. A frozen pretrained teacher yields SR@10 of 0.19 and the best ImgEdit among ablations (4.55).

VINCIE-7B scores SR@10=0.47 when each turn sees the full history including the clean source, and 0.09 when it sees only the previous output. Under that matched previous-turn protocol, all three MT-OPSD backbones score higher.

Why it matters

This is a cheap adaptation recipe: two thousand single-turn pairs, LoRA, no multi-turn labels. If a product already runs Qwen-Image-Edit, FireRed, or FLUX.2-klein in chained editing, this is a better fit than bolting on Emu Edit, FreqEdit, or VAE-LFA, especially when global restyles sit between local tweaks.

It does not close the gap to GPT-Image-2. What it shows is that single-turn skill can be transferred onto self-generated states, and that exposure bias here does not require multi-turn ground truth or RL to suppress. The pieces worth copying are identity rollout as a way to manufacture on-policy states, and matching teacher velocities on student-visited latents, which beats pseudo-label SFT by a wide margin.

Limitations

There is no Limitations section. The conclusion restates the contributions.

Both evaluation and teacher promotion use GPT-4o. The 23-session gate set is built with the same recipe as LME-Bench, only the sources are disjoint. Some of the metric lift may be the model learning to please that judge. There is no cross-model judge and no human study.

LME-Bench sources are all Z-Image-Turbo synthetics, not photographs. Identity rollouts only accumulate error from change-nothing turns. Error from real edits may look different. Gains at ten turns show some transfer; they do not show the error distribution was covered.

Even after MT-OPSD, the best open-source SR@10 is 0.52. Half the sessions still fail somewhere inside ten turns. GPT-Image-2 at 0.96 is another tier. VINCIE's native setting sees full history, so that comparison is not fully symmetric. Training runs at 512x512 while the benchmark sources are 1024x1024. Eight high-end GPUs are required. Wall-clock cost is unreported, and there is no test on real user photos.

Terms

• exposure bias: train on clean conditions, test on the model's own outputs, errors compound.
• identity rollout: repeat a leave-it-unchanged instruction to harvest the model's self-induced drift.
• on-policy self-distillation: train the student on states it sampled; the teacher is the same model under a richer condition.
• velocity matching: align flow-matching vector fields; here, at sparse query steps on the student's trajectory.
• SR@k / CR@k: fraction of sessions successful through turn k, and fraction already collapsed by turn k.

Source

• arXiv 2609.35611v1
• PDF
• Hugging Face
• GitHub · https://github.com/liangbingzhao/MT-OPSD

Related papers

• Best video deep-research report scores 63.3; fact coverage only 7.6%
• FilmFlow's four-agent planner passes 84.3% of blinded camera reviews
• Sensorimotor Alignment: Training-Free Metric Predicts Robot Policy Success, r=0.83 on Real Robots
• RealMe adapts 14B FLUX so eight 720p views reconstruct 2K volumetric video
• Tencent 4B VLM rejects absent objects at 67.4% while unifying boxes and mattes
• LayerEditor loses just 1% SSIM when the edit is rebuilt from its RGBA layers

All paper explainers