One task to train them all: teaching models to edit

Unified multimodal models usually need separate training for understanding, generating, and editing images, creating conflicts that hurt performance. This work reframes image editing as a general task that naturally demands both visual understanding and generation. The team created Uni-Edit-148k by converting visual question-answering data into complex editing instructions with reasoning logic, then showed that training on this single task alone improves all three capabilities on BAGEL and Janus-Pro benchmarks without auxiliary tricks.