AI's 'Imagic' that can edit one image with only text is too amazing

AI technologies that can edit images with text prompts are often talked about, but most of them have limited editing types or require multiple image inputs. However, ' Imagic ', which was announced on the preprint server arXiv on October 17, 2022, allows you to print a single image without having to manually specify the correction part of the image or input another image. It is possible to edit complex images with just a text prompt, so you can feel the awesomeness just by looking at the images published in the paper.

[2210.09276] Imagic: Text-Based Real Image Editing with Diffusion Models

'Imagic' utilizes a pre-trained Text to Image diffusion model, allowing editing of complex non-rigid bodies (objects that change shape). For example, you can output the image of the original bird (left) and the image of the bird that is about to take off with the instruction 'bird spreading its wings' (right) as follows.

Multiple targets are supported, such as 'two parrots kissing'.

It is OK to specify only one of multiple types of targets, such as 'a goat jumping over a cat'.

I also give instructions to edit the entire image in a large way, such as `` make the picture of the waterfall look like a child's picture ''.

You can see how flexible Imagic is when you look at the results of different directions on the same image. In the following, in addition to making the image of a dog standing on four legs sit and jump, we add a new object that specifically says ``make a toy in your mouth'', and multiple things like ``jump while holding a toy'' I combine the instructions of

You can also edit the shape and movement of people's hands, which AI tends to have trouble with.

It is also possible to output multiple results from the same instruction.

There are a wide variety of editing methods, such as changing the type of flower or changing the time period of the photo.

Imagic is published on GitHub and you can try it yourself.

diffusers/examples/imagic at main ShivamShrirao/diffusers GitHub

In addition, the results of the actual tests are explained below.

Understanding Imagic - Hoge Hoge

in Software, Posted by log1l_ks