Rarely does a finished image stop at a single edit. You add an object, adjust the lighting, then refine a detail — and the real difficulty is keeping the parts you were already happy with from being quietly destroyed by the next pass. Autoregressive diffusion models drift in exactly this way during multi-turn editing: textures blur little by little, the subject shifts slightly, and after a few turns the image is unrecognizable.

A recently released paper proposes a training paradigm aimed squarely at that failure mode. MT Forcing, from SenseTime Research with HKUST MMLab, Stanford and Tsinghua AIR, trains on the model's own multi-turn rollouts while drawing supervision from a single-turn teacher line — the teacher always starts from the original image, applies every accumulated instruction in one pass, and its output distribution corrects the student. The paper and project page are available at arXiv:2610.25404 and on the project site.

How the drift happens

The paper splits multi-turn editing failures into two categories: one is unstable visual quality, where artifacts accumulate across turns and details degrade; the other is inconsistent content, where regions that should be preserved and earlier edits get changed unintentionally. When a model continuously edits its own outputs, small errors compound step by step.

The framework builds on unified multimodal models (UMMs) with autoregressive diffusion. The authors describe two complementary training signals, one for each failure category.

Two signals: one holds quality, one holds consistency

Single-Turn Distribution Matching Distillation (Single-Turn DMD) targets visual quality. The student edits its own generated history one turn at a time, while the teacher starts from the original image, composes all instructions accumulated so far in a single pass, and provides distribution-level guidance. That way the teacher carries no error from one turn into the next, avoiding recursive error accumulation. The project page summarizes it as a single-turn teacher present at every turn.

Teacher-Forcing Guidance targets cross-turn content consistency. It supervises history-dependent transitions along real multi-turn trajectories: reference trajectories guide the warmup phase, and the signal continues as a regularizer during distillation. The authors state explicitly that per-turn distribution matching alone cannot guarantee cross-turn consistency — which is exactly why the second signal is needed.

Both signals act on the same student model, each covering one side of the problem.

The numbers across eight turns

The disclosed figures come from 100 eight-turn sessions in MTEdit-Bench, scored as eight-turn averages with SenseNova-U1.5 as the base model:

  • Instruction following (IF): 8.47 base → 9.66 with MT Forcing (out of 10)
  • Content consistency (CC): 7.09 → 8.40
  • Drift resistance (DRS): 3.83 → 4.30 (out of 5)
  • Matched-point displacement: 6.34 px → 0.74 px (lower is better)

The project page also spells out the evaluation protocol: eight-turn means weight every turn equally, and IF and CC follow the benchmark's failure-triggered scoring rule, where turns after a failure score zero. Pixel metrics measure mean displacement between matched points, with lower values indicating better spatial alignment, and each metric is interpreted separately.

Looking at turn 8 alone makes the gap starker: 7.68 against the base model's 3.22, a 4.46-point margin. The project page also provides a fifteen-turn editing comparison demo.

Reading it

The interesting part here is not any single score but a training-signal design problem that tends to get overlooked: in multi-turn settings, how the teacher is defined determines whether error compounds. Most methods assume teacher and student sit at the same turn, so the teacher climbs forward carrying its own accumulated error. MT Forcing instead sends the teacher back to the original image and folds all instructions into one pass, which in effect makes the supervision signal immune to historical error.

Two caveats are worth stating plainly. First, the repository has not released code yet — it is officially marked as coming soon, with training code, inference code and pretrained checkpoints all absent, so anyone wanting to reproduce immediately will have to wait. Second, the reported numbers are team-published results on MTEdit-Bench, the authors' own benchmark; third-party replication has not appeared so far, which is the discount one should apply when reading evaluation work of this kind.

Multi-turn editing is moving from demonstrable to deliverable — users will not accept an image that no longer looks like itself after five edits. Whoever trains "remembering what came before" into the model holds the actual entry point for interactive visual creation.