InstaDrag enables AI-powered automatic image editing through dragging

), but when I tried the demo myself, it was very slow. Recently, I noticed that NUS and ByteDance released a similar project called InstaDrag, which is said to be much faster in processing speed. Although the code hasn't been open-sourced yet and the demo can't run, we can still take a look at the principles outlined in the project and paper.

Effect

Single-Round Dragging

Multi-Round Dragging

Comparison

Speed enhancement

Principle

The INSTADRAG processing flow includes three components:

  1. : to ensure that the unoccluded regions remain unchanged;
  2. :Used to preserve the identity of the reference image;
  3. :Used to encode (operation point, target point) pairs.