Thankfully, there’s a fairly easy solution. You can’t print them, and you can’t zoom in far enough to crop out specific details without losing quality. The issue, of course, is that 1024 pixel images might work for the web, but if you need to use them for anything else - a real-world exhibition of AI art, for example - they’re way too small. Anything bigger than a few thousand pixels on a side is just too computationally expensive. Expanding the size of the output image even slightly results in an exponentially harder generation process.įor that reason, most AI image generation systems are limited to generating smaller images. The issue is that the more static you start with, the harder it is to get to the final result. In that way, a block of random pixels eventually becomes an astronaut riding a horse or anything else you can imagine. Systems like DALL-E, Midjourney, and their ilk create images via a process called latent diffusion.īasically, they take a big chunk of random static and slowly tweak it until it begins to look like the target image that a user had in mind.
Before we dive into the solution, let’s explore why DALL-E images are so tiny to begin with.
0 kommentar(er)
