Thank you for sharing this interesting work!
After reviewing the data and performance metrics, I’m particularly curious about UniWorld’s results on other editing benchmarks, such as GEdit-Bench used in Step1X-Edit. From what I can see, the ImgEdit benchmark appears to be in-domain with the training data (reference), so I wonder how Uniworld would perform on a more out-of-domain or challenging benchmark.

In my try, the out-of-domain editing is not satisfactory.

For text editing, the man's face is changed, prompt "write the text "gucci" to the bag":

Furthermore, the results of complicated prompt or motion manipulation are not satisfactory, too.