Tile-based image generation has been a long-standing problem in computer graphics. Traditional methods, such as ray tracing, are very computationally expensive and often produce low-quality images. A new method called stable diffusion is promising to solve this problem. The technique uses a diffusion algorithm to generate images from a set of tiles. The algorithm is stable because it does not change the output image when the input image changes. This makes it suitable for use in real time applications, such as video games or 3D rendering. The stability of the diffusion algorithm is one of its main advantages over other methods. Another advantage is that tile-based image generation can be done using standard graphics hardware, which makes it more accessible than other methods. The stable diffusion fork is currently being developed by researchers at Google Brain and UC Berkeley. It will be released as open source software later this year. ..
The completed version of the Stable Diffusion model can generate images with artificial intelligence right on your computer. There’s now a modified version that is perfect for creating repeating patterns.
“Tileable Stable Diffusion,” developed by Thomas Moore, is a modified version of the same Stable Diffusion model that can be used to generate incredible AI art and lifelike images. The main difference is that this fork (meaning software that has split off from another codebase) is designed to generate images that can be used in seamless repeating patterns.
Perhaps the best use for this is creating textures for objects in video games, especially since the grid size can be customized to match the requirements for a specific project or game engine. Patterns can also come in handy for device wallpapers, website backgrounds (especially if you’re going for a 90’s aesthetic), fill patterns in Photoshop, or even wallpaper for an actual real-life room.
According to Moore, the modified model “uses a circular convolution, so the model sees the image as if all the parallel edges were connected. Like a tube, but in both directions to form a torus.” The result is that each pattern doesn’t have a discernable edge when repeated in a tile pattern, just like game textures are supposed to work.
You can try out the model in your web browser to generate images, powered by Nvidia T4 graphics cards in remote servers, or you can run it locally on your PC with the code from the GitHub repository.
Source: @ReplicateHQ
