AI Clear performance analysis

So, there have been several threads about how good or bad AI Clear is, but most of those are based on a single noisy image with no point of comparison. I decided that I could do a series of images with different subjects to see how well it does.

Since the idea was born just now we will start with a boring indoor image (that’s potentially not perfectly in focus - but then that’s one of the things AIC does as well).

To make this “objective” I have decided to compare the Output of AI Clear against the camera itself - by stacking images to eliminate noise. That way we are getting a much cleaner image without introducing artifacts:

At this point in time Auto and Low basically fail the test right away. Without pixel peeping Medium seems quite good and High even better if we don’t look at the edges:

It seems like this has been introduced in the latest version - the more aggressive noise reduction also introduces some quite obvious artifacts. Additional noise, strangely enough, green/pink pixels on sharpened detail and faint auras. Also, if you look at the top of the “High” image it has introduced darker horizontal lines above the monitor and a big dark square inside the monitor.

I will do additional tests with different subject matter and lighting conditions, but so far it seems like this newer, faster version is doing worse than the initial one.

Edit: This is with the current at the time of writing Studio version of V1.12.6

Had a chance to do some more. This time with a more challenging subject, with some amount of detail.

This is a phone camera, with NR turned off, so even at ISO50 we have some noise. And once again I am using image stacking to acquire a “clean” image for comparison. It’s not perfect, but it is quite good.

On a surface level AI Clear has done really well, the noise in the sky is gone, the brick walls appear to have more detail, certain things appear sharper.

Now for a closer look:

This is where, again, I feel like the previous version did a better job. The new version, even or Auto or Low really likes introducing that Perlin noise style pattern and adds greenish/cyan pixels when trying to improve edges (with chromatic abberation).

On a more positive note let’s have a look at a noisy version of the same image at ISO800:

Quite impressive how true to the “real” colors the result is. However - I’m not expecting it to fish out real detail from all that noise, but it introduces artifacts (those splotches of sharper bits):

In general I have a question for Topaz - how do you acquire the image pairs of good/bad images? Artificially add noise to images? Take photos with different ISO values?

It seems like stacking would be nearly the best option (for photos without movement anyway) to get good image pairs for training the AI.