Recently, researchers from Google, together with the Max Planck Institute for Informatics and MIT CSAIL, released an image editing tool called DragGAN. This is a very intuitive image editing tool, users only need to control the pixels and direction in the image to deform the image.
It may not be intuitive to see such a description, why not take a look at the GIF picture below. In one sentence, DragGAN allows users to quickly adjust the position, posture, expression, size and angle of the subject of the photo.
Nowadays, there are many tools for converting text to images, but the prompt words you enter may not be exactly the same as the image you want in the end, and there may be some deviations in various aspects such as size and angle . The really interesting thing about DragGAN is that it can not only generate realistic images, but also control the images flexibly and precisely.
In contrast to many previous approaches, DragGAN researchers propose a general framework by not relying on domain-specific modeling or auxiliary networks. To achieve this, DragGAN consists of two main components: feature-based motion supervision, which guides handle points toward target locations; and a novel point tracking method, which utilizes discriminative generator features to localize handle point locations.
The researchers say their method outperforms SOTA in GAN-based operations and opens new directions for image editing. In the coming months, they hope to extend point-based editing to 3D generative models.
While the project's code won't be released until June, it currently has 4.8K Stars on GitHub .