Most image editors work well for a single edit but degrade quickly when asked to apply many successive edits: textures blur, prior edits drift, and preserved regions change unexpectedly. The core insight here is to train the editor on its own multi-turn rollouts while using a single-turn teacher to provide clean, cumulative-edit supervision—so the model learns to operate under realistic input drift without inheriting the teacher's errors.
Key Findings
- Sustained multi-turn quality: applying MT Forcing to a baseline UMM markedly improves multi-turn instruction following (eight-turn average rose from 8.47 to 9.66) and content consistency (from 7.09 to 8.40).
- Strong drift resistance: metrics report improved drift resistance (example: 4.30/5 vs. base 3.83) and far smaller matched-point displacement (0.74 px vs. base 6.34 px), indicating better spatial alignment across turns.
- Two complementary training signals: Single-Turn Distribution Matching Distillation (Single-Turn DMD) stabilizes per-turn visual quality by supervising student rollouts with a teacher that applies cumulative edits in one pass; Teacher-Forcing Guidance enforces history-aware transitions via reference trajectories during warmup and as a regularizer.
Who It's For and Trade-offs
- Great fit if you develop or fine-tune models for interactive, multi-step image editing workflows (e.g., chat-driven editors, iterative design tools) and need to maintain prior edits across many turns.
- Look elsewhere if you only require single-turn editing quality or cannot afford the extra training complexity: MT Forcing adds distillation and trajectory supervision steps and relies on a teacher model and multi-turn data to deliver its gains.
Method Snapshot
- Student trains on its own sequential rollouts to learn realistic distribution shifts encountered at inference time.
- Teacher is conditioned asymmetrically: it receives the original image plus a compound instruction (all edits up to the current turn) and generates a single-turn target; this avoids propagating student errors into supervision.
- Teacher-Forcing Guidance uses real or packed trajectories for warmup and regularization to strengthen cross-turn consistency beyond per-turn matching.
The approach is empirical and benchmark-oriented: it demonstrates notable numeric improvements on the MTEdit-Bench and comes packaged with model weights and demos for reproducibility.