This Website Uses AI to Enhance Low-Res Photos, CSI-Style
Let’s Enhance is a new free website that uses neural networks to upscale your photos in a way Photoshop can’t. It magically boosts and enhances your photo resolution like something straight out of CSI.
Once it receives your photo, the neural network goes to work, upscaling your photo by 4x, removing JPEG artifacts, and “hallucinating” missing details and textures into your upscale photo to make it look natural. You’ll need to wait a couple of minutes for the work to be done, but it’s worth the wait — the results we’ve seen are impressive.
We first tested the system with a press photo we had on hand from the Rylo camera that just launched. Here’s the original:
We then resized the image into a 500px wide photo.
Upsampling the 500px-wide photo in Photoshop to 2000px wide using the “Preserve Details (enlargement)” resample option produces a photo with horrible textures (look at the fingers):
But upsampling the 500px-wide photo using Let’s Enhance produces a much cleaner version of the image that magically restores realistic textures to the hands:
Here’s a crop comparison to help you more easily see the difference:
We did a number of other similar tests. Here are the results:
Cat
Face
Buildings
Face
The system is currently weak with things like eye realism, but it excels with things like hair, landscapes, and animals.
Let’s Enhance was founded by Alex Savsunenko and Vladislav Pranskevičius, a chemistry Ph.D. and a former CTO (respectively) who have been building the software over the past 2.5 months.
“We are researchers ourselves,” Savsunenko tells PetaPixel. “We took few state-of-art approaches, hacked around and rolled them into production-ready system. Basically we were inspired by SRGAN and EDSR papers.”
Let’s Enhance is currently in its first version and will continually be improved based on user needs and feedback. The current neural network “was trained on a very broad subset of images that included portraits at about 10% rate,” Savsunenko says. “The idea is to make separate networks for each ‘type’ of image. Detect the type uploaded under the hood and apply some appropriate network. The current version does better with animals and landscapes.”
Every time you upload a photo, three results are produced for you: the Anti-JPEG filter simply removes JPEG artifacts, the Boring filter does the upscaling while preserving existing details and edges, and the Magic filter draws and hallucinates new details into the photo that weren’t actually there before (using AI).
If you’d like to use Let’s Enhance on your own photos, head on over to the website and drag-and-drop an image into your browser.
