I tested a few desktop based machine learning methods to enlarge images recently.
There are paid online tools as well for this sort of thing but I’m not really interested in those because it requires subscription fees and signing up for yet another service online. Just my opinion though, maybe they are good too.
Overall I thought Rebelle 5 NanoPixel (only available in the pro version) was great for images if you add the texture of the paper to them as well but not as good with no texture added. On the other hand adding the texture and using Rebelle for images with no texture really made them look more natural than just using AI Photo & Art Enhancer alone. Both tools are useful in different ways for different usage cases.
I preferred Mediachance’s AI Photo & Art Enhancer for images that already have quite a bit of texture anyway (like a painterly brush stroke type). It works very well. I also like there tool because it gives several different modes to choose from and a content reframe tool which is useful for tweaking compositions.
Rebelle is better for preserving details from images that are more like drawings or smoother with fewer impasto / bristle like details than Mediachance’s software which is trained on DAP painting styles. It just so happens the RGBA painterly brushes in PD Howler and Krita work well with this.
Upscayl is the most basic of the ones I tried but it is free and open source. It did a good job of enlarging my squirrel image except it seemed to lose most of the colour information for the original image making it look more muted. Even trying to assign the correct colour space in Affinity Photo for the image did not resolve the issue whereas the image enlarged using AI Photo & Art Enhancer was fine in this regard.
Update 18th January 2023
AI Photo and Art Enhancer just updated with a new upscale setting called “digital art supreme”.
I tested this out on the Dark Crystal study (hand painted by me in Krita) again. I halved the size of the original painted resolution before I up scaled it on this setting as otherwise it would end up being way too huge. It still ended up being 8000px wide after the upscale.
It did a really good job of preserving the intended details and keeping the brush strokes looking nice. Before it smoothed out too much of the subtle brush stroke details and only emphasized the more impasto-ish brush strokes I had painted with the RGBA brushes. Definitely an improvement.
Update April 2023
I tried the updated version of Upscayl and it has been really improved from the version I tried before. It has several different preset modes for image enlargement built into it including: General Photo (Real-Esrgan), General Photo (re-macri), General Photo (ultramix balanced), General Photo (ultrasharp), Digital Art and Sharpen Image. The ultra mix balanced mode is pretty good and the other modes useful for different types of images. Bonus it is free and open source.
Use cases? Personally as someone who has done digital painting for a rather long time now, I’m thinking this technology may be useful for resizing smaller studies that I then want to paint in more detail without having to lose detail in resizing it in the old way. Of course it is also very handy if you need to make bigger resolutions for printing out art that may not have been large enough for certain uses if you didn’t paint it big enough to start with. Handy for people with lower end computers too as painting in huge resolutions is often way too impractical given the lag on brushes and memory use that may entail.
It could be very handy in game development, design and illustration to upscale textures and resources for art creation.
Another use of this technology is to help people who use AI art generation tools upscale their images to workable sizes since they are typically generated in very small resolutions.
Now I will not get into that too much at this point because personally I’m not using that technology right now, and I don’t write about things unless I’ve tried them myself as a rule. I’m going to see how the Law pans out when it comes to all that before I even consider using it. I wouldn’t feel comfortable using it right now given the concerns into ethics and copyright, which is something I’ve spent time learning about over many years and something I do respect. For instance I would never do anything commercial with fan art or studies of movie stills etc… those are for my own personal learning use only. Update April 2023 there are moves into ethical models for AI generative art and big corporations are getting into it as well. I feel like the technology is here to stay and I would be lying if I said I haven’t looked into how it could be useful to me.
Thanks for reading.