Multi-turn image editing extends the standard single-turn paradigm by letting users refine an image through a sequence of instructions. According to the paper's abstract, this sequential setup is better aligned with human-AI collaborative creation than one-shot editing. The authors argue that the key challenge lies in constructing high-quality multi-turn training data.

To address this, MTEdit introduces a training strategy based on packed trajectories, which the paper contrasts with rollout-based approaches. While the abstract is truncated and does not provide full experimental details, the title indicates that packed trajectories outperform rollouts for this task. The work thus points toward data representation and trajectory construction as central design choices for multi-turn editing models.