Using an API from Stability AI, you can zoom in on images. I tried it on my photo from the 90s – Živě.cz

artificial intelligence stability splurge new API for developerswhere they can enlarge the images.

The API uses AI models Real-ESRGAN And Stable Diffusion 4x Upscalerwhich is also available to the community, so with a powerful enough iron, you can scale images directly on your own device.

A library for both Python and the regular HTTP API

The API itself is available either through professional libraries Python And typescriptor by classic HTTP (REST). It is necessary to use it API keywhich you are You are born After registering on the DreamStudio platform.

Zoom in illustration using a neuron from Stability AI

Although the API charges for credits (from 1000 credits for $10), novice users get a certain potion for free for trials. Different actions burn different amounts of credits from dozens to units. He will tell you more This table.

I’m going to try to zoom in on my low quality photo from the 90’s

So let’s try Stability AI’s Image Upscaling API in action. This is how it turned out when I submitted my high-resolution image to the server 360 x 280 pixels And he gave the command to the automaton, Let it stretch to 2048 pixels!

Image source with a resolution of 360 x 280 pixels

Transformation on Stability AI servers, including connections and data downloads, took just three seconds, and the REST API returned the image as a Base64-encoded text string encapsulated in JSON.

Get a zoomed image using the HTTP API in a REST/HTTP client mailman

So I had to additionally translate it into a binary file. Python library will do everything for me.

Don’t expect miracles from pictures

You can see how it turned out in the photo below. Don’t expect miracles with pictures and such a powerful zoom. AI cannot yet account for the various structural data simply missing from the source. So with a little effort you can achieve the same thing in a graphic editor, but it will take a little more work.

After scaling by AI to 2048 x 1604 pixels

For comparison, I enlarged the original small image using the usual supersampling method with a slight sharpening in the Czech graphic editor Zoner Photo Studio X. Below you can see the clippings at 1:1.

Left: 1:1 cropping and zooming comparison using Stability AI and super regular testing with explanation

It will always be a problem with older, poor quality photos, however, with illustrations that are simpler in color, or with less aggressive enlargements of only tens of percent and fewer multiples, the results will be much nicer and usable.