Pixelated images can be the result of many factors, including personal design choices (e.g., pixel art). The idea of “depixeling” images and improving their quality isn’t exactly new, but the Face Depixelizer project takes a different approach. First, it harnesses the full power of artificial intelligence, and second... its results tend to be fabulously horrible. Would you like to see?
Beyond Waifu2x: The Quest to Depixelate
One of the main resources on the web for upscaling images without sacrificing (too much) quality is waifu2x. I personally use the offline version called Caffe, but it works the same way. At the end of the day, we’re “inventing” pixels and hiding noise while maintaining the best balance. However, nothing prevents us from interpreting pixelated photographs differently. For example, take very low-resolution or “censored” faces and try to create something new with the little information available.
Face Depixelizer: How It Works
That brings us to the Face Depixelizer project. When given a low-resolution image, the system searches and analyzes outputs from a generative model (StyleGAN) to generate a new “perceptually realistic” image at the same size as the original. The most interesting part is that the user isn’t required to use “human faces” in the process. The web discovered Face Depixelizer, and one of the first things people did was upload faces belonging to video game characters.
From Mario to Link, from Doom monsters to Duke Nukem and the Minecraft creeper, the results are strange and horrifying in equal measure. The algorithm clearly makes a huge effort to “place faces” where there aren’t any, and there’s a certain bias toward white-skinned people (look closely at the examples with Obama and Ali)... but for some reason I can’t stop thinking about the sun with a baby face from Teletubbies.
https://old.neoteo.com/selfie-2-waifu-retratos-en-version-anime/In the past few hours I’ve tried to do more tests on my own, but the system seems to be down. Running Face Depixelizer currently depends on the daily quotas of Google Drive, and with so much interest, it’s logical that the quota runs out quickly. Below you’ll find the link to GitHub, which in turn contains the link to Google Colab and the online version, but I imagine there will be a more complete and robust version soon. Until then, enjoy the results...
Official site: Click here
More Bizarre Results
