Super Resolution from Adobe...

Here are the crops, standard on left, multi-shot on right in the first image, vice versa in the second image (sorry, I deleted the screenshots and I am too lazy to redo them). Suffice it to say that if you see noise and obscured detail, it is not the multi-shot image.

In the first comparison, the shadow detail is obviously more well-preserved and useable in the multi-shot image. In the second comparison, you can see that detail, such as the mm ruled lines on the left edge of the Color Checker and the text on the white certification area of the Color Checker are much more clearly defined in the multi-shot image compared to the standard image.

Kirk

EDIT - Thanks to Capt Jack for pointing out that I reversed the description of which image was which in the first screenshot. His PM saved the day!

Here are screen caps, resized to 1600 px across, comparing the Leica multi-shot DNG to the ACR Super Resolution version of the single raw capture of the same scene (from the previous comparison). In ACR, one right-clicks on the standard raw image and then chooses "Enhance" - and choose super-resolution.

The Leica multi-shot image is on the left and the ACR Super Resolution is on the right in each comparison. The Super Resolution image contains noise and does not contain the level of detail in the true multi-shot image.

I did not boost the tone curves in these comparisons.

kirk

Same images, tone curve boosted.

And finally, a comparison of all three methods (original image, ACR Super Res of that original DNG, and the Leica Multi-Shot DNG). For the ACR Super Res and the Leica Multi-shot image, the rendering was reduced to 50% of its original size to match the pixel dimensions of the original raw file rendering. All three images were boosted to show the limits of the file. The screenshot was downsized from 1920 across (my display width) to 1600 to conform with the forum rules.

The Super-Res shot suffers from the same things as the original.

Congratulations if you made it this far into nerdom. Have a cold one and relax.

kirk

Those results make sense though don’t they? In one example you have multiple shots with different signals combined and in the other a single shot that is attempting to uprez the original image. Looks pretty good to me.

jcothron wrote in post #19209759
Those results make sense though don’t they? In one example you have multiple shots with different signals combined and in the other a single shot that is attempting to uprez the original image. Looks pretty good to me.

Correct - however, in the multi-shot image, you also are getting upres-ing.

kirk

That's a good demonstration and a good comparison at the end. Very clear!

kirkt wrote in post #19209763
Correct - however, in the multi-shot image, you also are getting upres-ing.

kirk

No doubt, it's great technology. In multi-shot images however, you have quite a few limitations (the scene can't move at all without issues for example) whereas with the Super Resolution that is a non-issue. I suspect the targeted use case for this tech (action photography subjects where the distance was too great and/or having the wrong lens on hand) aren't in a position to do multi-shot images. I shoot landscape and don't even try the multi-shot due to scene movement. (leaves blowing, water moving, etc.)

It's great tech for where it works, but it doesn't work in a LOT of practical use cases. Now for architecture, real estate, etc it would be perfect I would think.

There are definitely limitations to multi shot. Super resolution or gigapixelAI needs a good quality image to start with or else you will just get a larger, smeary version of your original.

The AI thing has become very fashionable in the last couple of years. It will be interesting to see if it is actually useful and becomes a part of people’s toolbox as AI models get more refined. One thing that is interesting too is the nature of the training sets of images and how the selection bias of those images ultimately affects your image. I remember receiving an email about a new version of some software and the marketing tag line was “Your image, our look.” Really? How about my look? Maybe I have an idea about what I’m trying to do with my image. It was sort of appalling.

Interesting stuff.

Kirk

kirkt wrote in post #19209929
There are definitely limitations to multi shot. Super resolution or gigapixelAI needs a good quality image to start with or else you will just get a larger, smeary version of your original.

The AI thing has become very fashionable in the last couple of years. It will be interesting to see if it is actually useful and becomes a part of people’s toolbox as AI models get more refined. One thing that is interesting too is the nature of the training sets of images and how the selection bias of those images ultimately affects your image. I remember receiving an email about a new version of some software and the marketing tag line was “Your image, our look.” Really? How about my look? Maybe I have an idea about what I’m trying to do with my image. It was sort of appalling.

Interesting stuff.

Kirk

Here's something to bend one's mind around. Using the Super Resolution on a multi-shot image. You start getting into some serious sizes that "should" be pretty darn clearn.


Regarding the baked-in looks, presets, etc. I've always tried to avoid them. I learned early on (Lr 1 or so) that anything I set up to happen automatically on import tended to stay that way outside of major global adjustments. As a result I customize nearly every setting for nearly every image. Having said that if I were a wedding photog my approach might be vastly different, or if I were in a field of photography where quantity of output mattered. With the type of work I do how fast I get an image processed doesn't apply.

super res on the multi-shot...

I have found that some of the Adobe AI filters in the PS experimental filters (Neural Filters) send the image to the mothership to process it. This might be the limitation here too. This way you don't need to ship a new AI model to the local end user like the Topaz products.

kirk

kirkt wrote in post #19209929
...

The AI thing has become very fashionable in the last couple of years. It will be interesting to see if it is actually useful and becomes a part of people’s toolbox as AI models get more refined. ...

I have it on good authority that the AI in Giga Pixel and Canon R series eye detect Auto Focus AI are both working very hard to eliminate the human race. They are very frustrated that they can not seem to manage it with the tools they have at hand. They are however in close communication with the AI in SawStop table saws like the one I have in the shop on campus, where they plan to at least begin eliminating human fingers,. for a start.

Saw a Youtube video of an M43 user using pixel shift in camera to make an 80MP raw file that was input to Super Res. The result was around 300MP. The results were surprisingly good.

kirkt wrote in post #19209929
There are definitely limitations to multi shot. Super resolution or gigapixelAI needs a good quality image to start with or else you will just get a larger, smeary version of your original.

The AI thing has become very fashionable in the last couple of years. It will be interesting to see if it is actually useful and becomes a part of people’s toolbox as AI models get more refined. One thing that is interesting too is the nature of the training sets of images and how the selection bias of those images ultimately affects your image. I remember receiving an email about a new version of some software and the marketing tag line was “Your image, our look.” Really? How about my look? Maybe I have an idea about what I’m trying to do with my image. It was sort of appalling.

Interesting stuff.

Kirk

Sounds, looks, and smells like Skylum - another example of a marketing department running a company and being totally out of touch with what the end user wants / needs.

What I find interesting with things like the AI marketing-driven craze and stuff like Microsoft (in windows) is the "you're too stupid, we'll do it for you" mentality. That mentality and the latest AI marketing fad/craze seem to be growing at the same pace.

CyberDyneSystems wrote in post #19210048
where they plan to at least begin eliminating human fingers,. for a start.

I can't LOL that because it is tragic when accidents happen. But it did bring back memories. Mr. Benz, my junior high school metal shop and wood shop teacher had 3 fingers missing on one hand. Yep, you guessed it - table saw.