Zero improvement - all model

I’m trying to upscale (2x) a smartphone (HEIC) image, the AI engine does nothing. I have a screen capture of the side-by-side preview for you to review.

Please provide the original image so that we can give it a try too!

I appreciate you sharing this image!

I was able to test this, and with the Low Res v2 model, it wasn’t making a difference for me either. Screenshot below.

However, I see changes when using the Art & CG model and Low Res v1.

Please try the other models and see how it reacts for you.

Maybe scale down the image before processing to bring the detail density closer to the pixel density. I made a downscale to 640 width and then an upscale by factor four.

Lowres V2:

Recovery Beta

I tried those two models and to me they’re worse than the original. Nothing like what you show on your website for photo improvement. I think this photo broke your AI engine!

so you want me to shrink the photo before enlarging it??? I don’t understand what will gain since the enlargement will be that much more.

the a. i. was trained with photos where the detail density is closer to the pixel density. scaling down images with a high pixel count but relatively low details before processing helps the a. i. to work on the image data the way it was trained for. :slight_smile:

Ah! Tried that but the image is worse, lots of artifacts.

how about this one:

There’s artifacts with that one too, look at the pink on the bed. AI engine needs improvement. :disappointed:

Yes I know. Even a. i. sometimes can’t do wonders. Or as an old saying reminds to us: garbage in, garbage out … :confused:

When I make a downscale to 300 pixel wide, that’s the smalles one possible with Photo AI and then scale it up by factor 6 via low quality model it looks like this:

Yes but Topaz Labs is spouting miracles with using the AI tool for enlarging images… just look at the website! Or are they all false images? :hushed:

You may need to reframe how you think of AI upscaling and advertising. The results shown on the website are as accessible to you as they are anyone else, as long as you start with the same original image.

When you see a demonstration of a product’s capabilities, one should generally assume that it’s an example of a best case scenario. It would not make sense for a company to do otherwise. When you go on a first date, you try to look your best and represent yourself in the best way possible, because you’re trying to make an impression. That does not mean that you wear expensive and painstakingly ironed shirts every day for every activity, and it would be unreasonable for your date to assume that your appearance was representative of your normal.

Advertising is advertising whether for a person or software.

Additionally, the nature of AI upscaling is such that results can vary widely based on characteristics of an image that are not obvious or even visible to a human looking at the image, because 1) a computer is not a human and doesn’t have eyes, and you cannot transpose your experience of “seeing” as a human onto that which neither sees nor is human; and 2) the exact “thought process” of AI is opaque at best.

Your subjective assessment of the quality of one result is not indicative of any other result. You cannot look at the “best of the best” examples of results, and then have the expectation that every single picture in existence, upscaled with every possible configuration of settings, will always result in an identical subjective impression of quality.

Your current perspective will serve as your own personal disappointment factory, but you have the ability to get the factory greatly downsized so it’s not such an eyesore from your porch.

I wasn’t expecting the same results as the advertised images but I was expecting some improvement not the opposite. I think my factory is justified. I agree with you that the factory should not be an eye sore from the front porch assuming that the date can deliver at least in a few of the promises.