Despite tech giants rushing to incorporate artificial intelligence into their products, AI features usually leave a lot to be desired – and it’s gotten to the point where I’d just groan every time a new app developer or hardware manufacturer touts LLM-powered facilities as the next big thing. That said, there are a couple of examples where AI tools can be surprisingly effective, with NotebookLM being a prime example of how artificial intelligence can simplify complex tasks.
Better yet, there’s a lot of cool stuff you could do with LLMs if you’re into self-hosting. Take local image upscalers, for instance. As someone with gigabytes of old, low-resolution photos stashed on a NAS, local AI models let me enhance their fidelity without any strings attached.
What’s the point of running an upscaler on a local machine?
Better privacy, and zero paywalled features
Although there are plenty of platforms that can let me increase the resolution of an image, I’d rather not touch them with a ten-foot stick. Privacy is always a must for me, and I especially don’t want some third-party firm gaining access to my personal images under the guise of upscaling them.
Then there’s the fact that the majority of these websites have paywalled features. And I don’t just mean advanced scaling methods, either. When I attempted to up the resolution on some stock images, some platforms would add an annoying watermark that could only be removed by shelling out money.
Not to mention, most AI upscaling tools have restrictions on the number of files I can upload to them, and increasing the cap requires – you guessed it – spending more cash on tokens. Throw in long waiting times, and you’ll notice why I prefer a locally-hosted upscaler that neither voids my privacy nor forces me to spend extra bucks on virtual tokens.
I don't want to rely on Photoshop, either
If you’ve used Photoshop in the past, you may be aware that the app has a couple of methods to upscale images, including a generative upscaling utility added last week. Technically, it’s better than most online platforms I’ve tinkered with, but considering Adobe’s anti-consumer actions in the past, I’m not a big fan of using Photoshop to upscale private images.
Setting up ComfyUI
And selecting the right upscaling model
My game plan was to find a model – alongside a tool that runs it – that’s light enough to run on consumer hardware without requiring an RTX 5090 card and doesn’t force me to install obscure libraries or drivers. ComfyUI is perfect for the task, and since the tool is available for Windows, I could easily use it on my daily driver.
For reference, my PC houses an RTX 3080 Ti, so I’ve got 12GB VRAM to work with. The 4xNomos8kDAT model on OpenModelDB seemed great for the task, as it works well with landscape shots (which form a bulk of my old photos) without requiring obscene amounts of VRAM.
First, I downloaded the ComfyUI installer from the official website. After running it with Admin privileges, the wizard showed the “We were unable to find git on this device” pop-up, and since I’ve relegated a separate VM to my development tasks, it was time to set up Git on my main PC.
So, I downloaded the Git installer, ran it as an administrator, and went through a barrage of menus. In most cases, I went with the default options, and once Git had finished installing, I switched back to the ComfyUI installer.
Once the installation wizard detected Git, it asked me to choose the drivers for my GPU, and I went with the default option yet again. After selecting the Download Location, disabling the Migration options, and unchecking the toggle to send usage metrics to the ComfyUI, the installer began setting up cpython, pytorch, and other packages. Before long, the installation process was finished, and it was time to tinker with some AI workflows.
Configuring the workflow
It’s a cakewalk
Since ComfyUI uses a node-based format to create workflow chains, it may seem a bit complicated at first glance. However, the official ComfyUI guide includes a basic workflow to upscale photos, and importing it is as simple as dragging the tutorial image into the app’s dashboard.
However, I still had to configure ComfyUI to use 4xNomos8kDAT. I downloaded the model from its OpenModelDB link and copied it to the upscaling_models folder within the ComfyUI directory. After restarting ComfyUI, I was able to select 4xNomos8kDAT under the model_name section of the Local Upscale Model node.
Just to test things out, I uploaded some low-resolution images to the Load Image node and hit the Run button. Well, the results were surprisingly impressive for 720p images captured using a really cheap camera.
There are tons of cool models to experiment with
So far, I’ve only tried a handful of models from OpenModelDB. 4xNomos8kDAT works pretty well, though the older RealESRGAN_x4Plus is a few seconds faster with a slight decrease in quality. For portraits, I’ve had decent luck with 4xFaceUpDAT and Remacri. I’ve only conducted a few tests for illustrations, but AnimeSharpV4 provided satisfactory results.