MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion

MT Forcing. Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion. By Houyuan Chen and eleven others, from SenseTime Research, the Hong Kong University of Science and Technology, and two other institutions. Published on the eighteenth of October. Its code repository is worked out: about fifty-five source files, with a readme.

Multi-turn image editing changes a picture according to a sequence of instructions, one after another, which is how people actually work when they create visual content interactively. The trouble is that existing methods drift after only a few turns, and the drift shows up in two separate ways. The first is unstable visual quality: artifacts accumulate and details degrade. The second is inconsistent content: regions that were supposed to be preserved change anyway, and edits made in earlier turns are quietly undone or altered.

MT Forcing is a training paradigm aimed at both failures, built for unified multimodal models that rely on autoregressive diffusion. It pairs one form of supervision for each kind of drift. Visual quality gets a teacher that never has to edit recursively. Content consistency gets guidance along reference trajectories that carry the history of the edits.

The first piece is Single-Turn Distribution Matching Distillation, or Single-Turn DMD. The student produces its own rollouts across turns, and those rollouts are supervised by a teacher that applies all the cumulative edits to the source image in a single turn. Because the teacher reaches the target in one step, its guidance on visual quality is free of the recursive error accumulation that the student is prone to.

The second piece is Teacher-Forcing Guidance. It supervises the history-dependent transitions, the steps from one turn to the next, on reference trajectories. It is used during a warmup phase, and again as a regularizer during distillation. Its purpose is cross-turn content consistency, which per-turn distribution matching alone cannot guarantee.

Extensive experiments show that the approach significantly improves the multi-turn editing capabilities of unified multimodal models. The gains land on exactly the two axes the paper set out to fix: more stable visual quality across turns, and stronger consistency of content from one turn to the next.

A long editing session fails for two different reasons, and MT Forcing answers each with its own signal. One teacher sees the whole edit at once, so quality never inherits compounded errors. A second form of guidance watches the history, so what was settled stays settled.