By feeding the original polygonal images from a Sega Saturn emulator into an image-to-image mode of the Stable Diffusion text-to-image deep learning model, artist Colin Williamson was able to turn 1993’s Virtua Fighter characters from polygonal people into much more realistic versions of themselves. We sure have come a long way, haven’t we? I still remember having a crush on the incredibly polygonal 1996 version of Tomb Raider’s Lara Croft.
In addition to feeding the original images into the deep learning model, Colin also added text prompts attempting to describe the style of the characters and their clothing. Once he had a prompt that was producing the desired results, he’d then run batches of the images through the AI. These are some of the results.
Are the AI-generated versions better than the originals? Well, they’re certainly less blocky and more realistic. But I’m afraid for strictly scientific purposes, I’m going to have to request that you run 1996 Lara Croft run through the model. Also, if you could add the prompts ‘dream girl’ and ‘Sports Illustrated Swimsuit Issue,’ that would be great.
[via arstechnica]
0 Commentaires