How poorly upscale AI really works
Answered
And here quite practical:
Here is an extreme example in a 400% view. Die Original size: 768X1024, enlarged 4x.
The first image is the original.
As you can see very nicely, the mouth and nose are barely visible. Extreme artifacts.
The second image is the result of Luminar Upsclae AI.
Again you can't see any difference to the original.
The third image is the result of Topaz Photo AI.
Here you don't need to lose any words. It is so much better.
I tried this on quite a few images in Canon RAW CR2, jpeg, and TIFF formats. And didn't matter if it was landscapes, architecture, or faces. The result was always the same. I found these results with all images without a single exception.
While NEO upscale AI only upscales the pixels, does not restore any details, and thus ultimately does nothing different than any program that performs pixel interpolation to enlarge, Topaz Photo AI establishes, in addition to pixel number enlargement, mainly restore details. It is enormous what the artificial intelligence behind Topaz Photo AI achieves.
And that dear Skylum team, is the bar, no less!
-
Yes, amazing result but I believe that it has been achieved by reconstructing the face on the basis of a database of images rather than by restoring details of the picture itself.. So, this is not the face of the person in the original photo but rather a 'lookalike'. And I've seen some TPai results that were not very good.
-
Hi Hilltops,
thank for the feedback.
Topaz does not simply insert a new mouth or nose here from other images. Rather, it reconstructs lost details from pixel-level details and the actual details of a mouth and nose (from collected data from millions of images). AI is doing here what our brains do. Just maybe not as well. But amazing what comes out. The image was a scan of an old analog image. I scanned it in low quality on purpose to see what the software would do with it. And the result is so close to the original that it's almost indistinguishable.
I also tested other images with more resolution, taken from a digital camera. Again, the result was amazing. With architecture or with landscapes, many details became visible again.
What you have to remember is that there is more to see on the pixel level than our eye can capture. Pixel level and the associated binary data can all be processed mathematically. And the more accurate the mathematical model used for this purpose, the better the result. An AI uses learned data from millions of other data (images) as a reference to establish order from a pixel mess.
-
"Rather, it reconstructs lost details from pixel-level details and the actual details of a mouth and nose (from collected data from millions of images)." That's exactly the point I was making. Of course Topaz uses the structure of the original photo as starting point, hence the 'lookalike' I was talking about.
To be sure, I'm very happy that such technology is available and getting better and better. However, it's good to know what the underlying process entails. That's why I posted earlier.
-
Hi Hilltops,
you are right here, of course. I had misunderstood you a bit in your post...sorry.
I just expected more from upscale AI, because the way it currently works is pointless. Because with any program I can enlarge an image with a pixel interpolation. And that's all upscale does. Actually, it's worse, because upscale causes some ugly artifacts, which are shown excellently on YouTube by Anthony Turnham, for example. He also compares with Topaz and finds a clear verdict here.Here is the link:
https://youtu.be/AXdmj8U0fwI
Please sign in to leave a comment.
Comments
6 comments