MTEdit dataset trains multi-turn image editing without rollout drift
Original title: MTEdit: Packed Trajectories Beat Rollouts for Multi-Turn Image Editing
A 500K-trajectory dataset and MTEdit-Bench back a fine-tune that leads evaluated open-source editors.
Researchers introduce MTEdit, a multi-turn image editing dataset built with packed trajectory canvas generation so multiple edit states are produced in one visual context. The set covers step-wise and context-aware editing, with 500K trajectories of up to 8 turns and an average of 5.38 turns. Fine-tuning Qwen-Image-Edit-Causal yields Qwen-Image-MTEdit, which the authors say is the best open-source model on their new MTEdit-Bench.
- MTEdit has 500K trajectories, up to 8 turns, averaging 5.38 turns.
- Packed trajectory canvases jointly generate edit states to cut visual-quality drift.
- MTEdit-Bench scores instruction following, consistency, drift resistance, and history use.
- Qwen-Image-MTEdit is the fine-tune of Qwen-Image-Edit-Causal.
- Authors say it beats other evaluated open-source editors on the bench.
From the original article
1 min read at huggingface.co
Multi-turn image editing aims to modify an image according to a sequence of instructions, making it better aligned with human-AI collaborative creation than single-turn editing. However, constructing high-quality multi-turn training data remains challenging, as conventional sequential rollout accumulates errors and introduces visual-quality drift across editing turns.
By Houyuan Chen, Fanzhou Wang, Xiangyu Fan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.