Reproduce “MTEdit: Packed Trajectories Beat Rollouts for Multi-Turn Image Editing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Houyuan Chen, Fanzhou Wang, Xiangyu Fan, Songlin Yang, Zhitao Yang, Lvmin Zhang, Pei Yang, Hao Zhao and 3 more
MTEdit constructs multi-turn image-editing data by generating several editing states together on a packed trajectory canvas, then fine-tunes an open-source editor on the resulting dataset. In a matched eight-turn comparison, the PTC-trained model scored 4.22 on Turn 8 drift robustness, versus 3.96 for sequential-rollout training. This supports an advantage in long-chain editing under the tested setup, not elimination of drift.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.