This is something that may or may not be more suitable for another project, but I wanted to bring this to everyone's attention, especially as I've been running into issues recently that this could easily solve.
OpenAI recently published papers showing a new technique for image completion. You feed the AI an incomplete image and it returns a completed image. For instance, you can give it half of a picture of a cat, and it will return several possible complete images of the cat. The technology is amazing, and just describing it doesn't do it justice. Here is a video explaining how it works with examples.
So how is this relevant to waifu2x? Well, one big issue I often run into is when I try to resize images. waifu2x does a fantastic job of making my images larger, but what about when I want to change the aspect ratio of an image?
Let's say, for example, if I have an 800x600 image and want to make it into a 1920x1080 wallpaper. I can use waifu2x to enlarge it to 1440x1080, but that's the wrong size. So my only options are enlarging it more and cropping it, or filling the borders with solid color, or a blurred copy of the image, or something of that sort which is both non-ideal and labor-intensive. But with image-gpt, I could use an AI to automatically fill in the rest of the image for me.
Now of course, there are serious questions of implementation and training - I don't know exactly what model image-gpt is trained from, but I have a feeling it may not be ideal for anime-style images. So I can't say as to how feasible this may be. But now that we have some great tools for working with anime-style images such as image enlargement (waifu2x), lineart colorization, and decensoring, this may be a fantastic next step.
The landing page for image-gpt is here and the GitHub page is here.
This is something that may or may not be more suitable for another project, but I wanted to bring this to everyone's attention, especially as I've been running into issues recently that this could easily solve.
OpenAI recently published papers showing a new technique for image completion. You feed the AI an incomplete image and it returns a completed image. For instance, you can give it half of a picture of a cat, and it will return several possible complete images of the cat. The technology is amazing, and just describing it doesn't do it justice. Here is a video explaining how it works with examples.
So how is this relevant to waifu2x? Well, one big issue I often run into is when I try to resize images. waifu2x does a fantastic job of making my images larger, but what about when I want to change the aspect ratio of an image?
Let's say, for example, if I have an 800x600 image and want to make it into a 1920x1080 wallpaper. I can use waifu2x to enlarge it to 1440x1080, but that's the wrong size. So my only options are enlarging it more and cropping it, or filling the borders with solid color, or a blurred copy of the image, or something of that sort which is both non-ideal and labor-intensive. But with image-gpt, I could use an AI to automatically fill in the rest of the image for me.
Now of course, there are serious questions of implementation and training - I don't know exactly what model image-gpt is trained from, but I have a feeling it may not be ideal for anime-style images. So I can't say as to how feasible this may be. But now that we have some great tools for working with anime-style images such as image enlargement (waifu2x), lineart colorization, and decensoring, this may be a fantastic next step.
The landing page for image-gpt is here and the GitHub page is here.