MTEdit
A large-scale dataset for multi-turn image editing training that uses a Packed Trajectory Canvas (PTC) pipeline to jointly generate multi-turn editing states. It contains 500,000 trajectories and supports up to 8 editing turns.
Timeline (last 90 days)
The MTEdit dataset contains 500,000 multi-turn image editing trajectories, supporting up to 8 turns, with an average of 5.38 turns per trajectory.
The MTEdit dataset covers two types of scenarios: stepwise editing and context-aware editing.
The core innovation of MTEdit is the Packed Trajectory Canvas (PTC) generation pipeline, which generates multiple editing states simultaneously within a unified visual context, thereby reducing visual quality drift across turns.
The research team used MTEdit to fine-tune Qwen-Image-Edit-Causal, obtaining Qwen-Image-MTEdit.
The MTEdit dataset is currently one of the largest multi-turn image editing datasets.
All Facts (5)
The MTEdit dataset contains 500,000 multi-turn image editing trajectories, supporting up to 8 turns, with an average of 5.38 turns per trajectory.
50%UnverifiedThe MTEdit dataset covers two types of scenarios: stepwise editing and context-aware editing.
50%UnverifiedThe core innovation of MTEdit is the Packed Trajectory Canvas (PTC) generation pipeline, which generates multiple editing states simultaneously within a unified visual context, thereby reducing visual quality drift across turns.
50%UnverifiedThe research team used MTEdit to fine-tune Qwen-Image-Edit-Causal, obtaining Qwen-Image-MTEdit.
50%UnverifiedThe MTEdit dataset is currently one of the largest multi-turn image editing datasets.
50%