Throughout my artistic career, I had to prepare my images for printing probably thousands of times. Preparing photographs before sending them to magazines, for a gallery show, or for a fine-art print sale. But in photography, you rarely face the necessity to upscale an image. Or to upscale an image that has indistinct details and a lot of... uncertain spots. :) And most definitely, it is the rarest event to upscale a 1024x1024 pixel image for high-quality printing.

Yet here we are, with Midjourney's beautiful but elusive results that, in V4, can only go as big as 1024x1536 pixels. What is the best way to upscale them and prepare them for printing? In this study, I will share experimentations and observations that led me to my favorite upscaler for AI-generated art.

There are many ways to upscale an image, and it's not this study's goal to cover all of them. I will, however, try some of the most popular options, both free and paid.

After considering each option's speed, quality, and user experience, I settled on using Gigapixel AI. Despite a hefty 99,99$ price tag, my time and mental health seemed more valuable. :)

Someone with a stronger will, more dedication, time, and a technical mindset could squeeze better results out of some of the methods in this study. For instance, selecting a correct Upscaler model can drastically affect the upscaling outcome. But going through lists of literally hundreds of them, let alone downloading, installing, and waiting between ten to ninety minutes to see each result—that's a task for a dedicated research institute. :) Instead, I went with the most common models and (mostly) default values in most cases.

My upscaler of choice for many good reasons—Gigapixel AI—is an all-in-one AI-based upscaler with a buy-once-own-forever license. When you launch the app, it offers to download the most recent upscaler models. Afterward, it will regularly download updates and occasionally add new models.

Gigapixel AI offers a minimalist and clean interface. Every control is in its place, and the overall experience is straightforward and intuitive.

You drag and drop your images to Gigapixel AI's window. They appear as a list with parameters for each position (that can be changed individually or for all images simultaneously).

A handy Auto mode lets Gigapixel AI decide which model best applies to your images and with what parameters. Or you can set everything manually. That includes Suppress noise, Remove blur, Fix compression, Face Recovery, and Gamma Correction.

To help you with these choices, the central part of the screen is occupied by a before-after comparison preview. You can zoom in and out and drag the loupe across your image. The preview only runs the selected upscaler model on the zoomed part of the image, so it works faster than complete upscale—to show how this particular fragment will look after you launch the upscaler.

Finally, you choose where and how you want the upscales saved and hit Save images. Voìla, done!

In this comparison, Gigapixel AI upscales are put against original images resized to have the same resolution. In all the following tests, I will be comparing the results from the upscaler in question to those of Gigapixel AI.

Also, I won't be posting all results for each upscaler. You can download all fragments via the link at the end of this study.

Gigapixel offers several built-in AI models (see specs below). However, I found the Standard mode the most useful, delivering the best results in each benchmark position. Here is a comparison of Lines vs Art & CG vs Standard models with Line Art test.

Super Zoom is a neural filter in Adobe Photoshop that uses artificial intelligence to—as its name suggests—increase the resolution of an image. The feature uses deep learning to analyze the details of an image and generate new pixels for smooth upscaling.

SuperZoom has an additional option called "Enhance face details." Let's check what it does: we have a benchmark exactly for that.

A decent enhancement, especially in the hair! Let's see how Super Zoom with Face enhancement compares to Gigapixel AI with Standard model.

Google Colab is a free online platform that allows users to write and run code in a web browser. It also provides access to powerful computing resources, such as GPUs, for running complex machine-learning models. Numerous colabs can upscale images (and not only that).

For this test, I chose SuperRes Diffusion—batch upscaling and super-resolution colab based on Latent-Diffusion. If you decide to try it yourself, I recommend this short and comprehensive article on how to use it.

chaiNNer is "a flowchart/node-based image processing GUI aimed at making chaining image processing tasks (especially upscaling done by neural networks) easy, intuitive, and customizable."

It is easy and intuitive—after a particular learning curve. ;)

chaiNNer is a powerful tool to do so many things that upscaling becomes a fraction of its functionality. It is a spaceship! This is fantastic if you're into space traveling but overwhelming when you need a drive to a local supermarket. 8)

After downloading and installing chaiNNer, you will need to also download and set up the libraries you want to use. Upscale.wiki offers a massive list! There are hundreds of them: from universal-purposed models to very specifically-targeted ones (VHS restoration, Super Mario textures upscaling, models trained on coins, or cats, etc.). They have descriptions, but not all have examples of their outcomes. So you might have to choose from a few dozen of seemingly similar entries.

chaiNNer's interface is nicely designed, but it might scare an unprepared beginner. Especially if this is your first time working with node-based software. The principle, however, is pretty simple. There is a canvas where you arrange cards that represent inputs (like your image), actions (like applying an upscaler model), and outputs (like saving the result). By setting/dragging links between those cards, you define the workflow.

With chaiNNer I ran all tests using three different models: 4x Valar (for realistic photos), Face-Ality V1 (trained on faces), and Digipaint (an upscaler for digital art).

Upscayl is a free, open-source desktop application that works on macOS, Windows, and Linux.

With its lightweight interface (maybe not the best one aesthetically ]:->), Upscayl offers a straightforward and fast user experience while utilizing advanced AI models and producing a decent result.

For the final part of this test, I decided to see how Gigapixel AI upscales will behave on paper—with a particular focus on fine art prints. Each sample was printed at 300DPI on A3 size (297 mm × 420 mm, or 11.7 in × 16.5 in.)

Gigapixel AI showed excellent results even on the most demanding paper. All prints turned out perfectly detailed, without artifacts or pixelation. A3 is a perfect size for both selling prints directly or exhibiting them at a gallery show (or on your wall!). And Gigapixel upscales absolutely shined as A3 pictures. However, my impression is that you can go at least 1,5x bigger without losing quality.

In this study, I tried to show specific tools but also outline the ways you can go if you are serious about upscaling your images. Whether you are ready to buy yourself a headache-free ready-made solution, or you will settle with a simpler free tool, or you will go into node-based programming and simply coding—there are plenty of options that await you on all those paths.

Happy midjourneys!

— Andrei

P.S. A very special thank you to Phil Akhnazarov for his help with overcoming chaiNNer! <3