Testing Some Generative AI Image Upscalers

I'm working on a new census web page, and have run into a situation where the only image I have of an item is so tiny, that it is relatively useless from a detail perspective. It's only 158 pixels wide by 227 pixels tall:



So I thought this was a good opportunity to test the burgeoning number of "free" generative AI upscalers to see if any of them are worth a hoot compared to just doing an upscale in Photoshop.

Note: Were I to use an AI-generated image on my web page, I would do two things: 1. The image would be clearly marked as AI-generated, and 2. the original source image would also be provided.

Some of these results may be improvable once you go past the free tier to a paid plan, and/or digging deep into the settings/parameters, but I wanted to test what they can do "quick & dirty". So no paid options, and only using preconfigured settings sets.

Don Denman (R.I.P.) had touted Topaz Gigapixel, and while they are still a contender in the space, they have unfortunately changed their pricing model. Instead of a one-time license purchase, it is now solely a subscription-based model, and an expensive one... unless you sail the seven seas.

Most of the results are pretty awful, either blowing out the background details or making insane leaps to "fill in the gaps". What you end up with are stamps that look like they were painted by Salvador Dali... in other words, not viable in the least.

Also look at the difference in colors produced by the different upscalers.


1. Original image.



2. Adobe Photoshop built-in upsampling.



3. Adobe's Generative AI.



4. https://ai.nero.com/image-upscaler



5. https://upscayl.org/ (digital art settings)



6. https://upscayl.org/ (standard settings)



7. https://upscayl.org/ (ultrasharp settings)



8. https://upscayl.org/ (high fidelity settings)



9. https://imgupscaler.ai/ (beta setting)



10. https://imgupscaler.ai/ (business setting)



11. Topaz Labs https://app.topazlabs.com/enhance/upscale (low-res setting)



12. PixelBin https://upscale.media/upload




All of the above are completely unacceptable. However, there is one that IMO is surprisingly good given the data it had to work with. It still has issues (e.g., look at the mouth of the portrait in the top stamp vs. the bottom), but IMO the results are far superior to the above.

13. https://imgupscaler.ai/ (V2 Pro setting)




Compare this to an actual high-res scan (different multiple, as I don't have access to the one this thread is about). When compared next to one another, the mistakes jump out. The background ornaments aren't quite right and the microprinting under the portrait and the BEP imprint don't come through... which isn't surprising given the resolution of the original.



Side-by-side:




Far from flawless, but I'm actually impressed at what ImgUpscaler could pull out of the original image. A few generations down the line and who knows what will happen, especially if the AI pulls in context from online image data in addition to what can be gleaned from the source image.

Until then a flood of questionable images which will break trust of any image will result.

No different from what AI is currently doing to news images, audio, and media, and the havoc it is wreaking in academic fields, not to mention consumption of resources (energy, water, RAM, etc.).

In hopes of increased automation and tech progress; let's be honest: shareholder profits) We've opened a Pandora's Box and we've only seen the tip of the iceberg of both the good and bad that can come of it. It's also an area virtually impossible to regulate.

This is reminiscent of the differences observed between reprints of local stamps and the originals.

I can't unsee the one with tongues for eyes...

10. https://imgupscaler.ai/ (business setting)

Nor does it make it any more right, moral, correct or actually useful.

Time to go old school and have everyone hand write answers to a one on one oral test. If one can pass that, one will have a chance to recognize the errors of AI. Otherwise, just call out next.

AI is still governed by GIGO, just a more highly polished lump of excrement results. And such a system is labeled proprietary to avoid transparency and revealing built in bias and errors.



Edit: Here is an example plucked for today's headlines: https://www.sfchronicle.com/tech/ar...21221662.php




There even the judge used AI and came up with a faulty court order.

I usually use the Resample option in IrfanView for this kind of problem. Works and no AI required :-)

I'd be curious to see what your results are using Irfanview on the original image above (I don't have that software installed). Unless it is significantly different from Photoshop's upsampling, it wouldn't be viable IMO.

Your original image IrfanView resized 400% He looks like the same person in both stamps here.

Thank you. Looks virtually identical to the photoshop upsampling.

I'm only 51; but maybe a grumpy old man already (judge for yourself!) But why not just a scan?

Freely admit I think AI is going to not make society as a whole better, to wax philosophical for a moment; but sometimes what works just works.

Bravo on the numerous samples though, you really put a lot of time in to that! But some of those examples look really creepy. Yikes!

Because I'm compiling a census and all I have available is that low-res image. The whereabouts of the original item are unknown.