MTEdit: Packed Trajectories Beat Rollouts for Multi-Turn Image Editing

MTEdit. Packed Trajectories Beat Rollouts for Multi-Turn Image Editing. By Houyuan Chen and nine others, from SenseTime Research, the Hong Kong University of Science and Technology, and two other institutions. Published on the sixteenth of October. No code repository is linked to it.

Multi-turn image editing modifies an image according to a sequence of instructions, one after another, rather than a single request. That fits the way people and AI create things together better than single-turn editing does, since collaboration rarely ends after one instruction. The trouble is training data. High-quality multi-turn examples are hard to construct, and MTEdit is a new dataset built to supply them at scale.

It covers two editing scenarios: step-wise editing, and context-aware editing. In total MTEdit holds five hundred thousand multi-turn editing trajectories. Each trajectory runs up to eight turns, and the average is five point three eight turns per trajectory. By the authors' account, that makes it one of the most extensive datasets for multi-turn image editing.

How it was built is the heart of the contribution. The conventional approach is a sequential rollout: make an edit, then edit the result, then edit that, and so on down the chain. Every turn inherits whatever went wrong in the turn before, so errors accumulate and visual quality drifts as the turns pile up. MTEdit's pipeline is centered instead on what the authors call packed trajectory canvas generation. Several editing states are generated jointly, inside one unified visual context, instead of one at a time. That joint generation reduces the drift across turns. It is the comparison the title announces: packed trajectories against rollouts.

For those evaluating editing models, there is MTEdit-Bench. It tests both scenarios, and it scores four abilities: following editing instructions, keeping content consistent, resisting visual-quality drift, and making effective use of the history of earlier turns. For those training models, the dataset itself is the offer. As a demonstration, the authors fine-tune Qwen-Image-Edit-Causal on MTEdit to obtain Qwen-Image-MTEdit. They then evaluate a broad range of closed-source and open-source editors on MTEdit-Bench, including their own. Qwen-Image-MTEdit achieves the best performance among the open-source models evaluated.

Where the drift comes from decides how to remove it. If errors compound because each turn is built on the last, then generating the turns together in one canvas sidesteps the compounding. The fine-tuned model is offered as the evidence that data built this way improves multi-turn editing in practice.