I came across this video (a bit long for my taste) that revisits the eternal debate between Full Frame and cropped formats, and the alleged loss of quality for us "normal" enthusiasts. Here, the comparison is even more extreme because we are talking about an excellent but older M43 camera as the Panasonic G9, medium format digital cameras like the Fujifilm GFX 50S, and medium format film cameras.

The guy evaluates the image quality by looking at digital files and physical prints measuring 16x20 inches. Because the digital files were cropped to a square ratio to match the film camera, the magnification level of the prints equals a larger 26x20 inch format.

During the evaluation, the MF digital camera displays finer details and smoother rendering. On close inspection of the physical prints, the medium format images show more distinct lettering and sharper elements. When observing the prints from a normal viewing distance, the results from the M43 and the MF cameras appear identical. It is necessary to stand very close and examine the prints side by side to notice the variations. Without having two prints next to each other, it is difficult to identify which camera system captured the image.

The video also notes that using AI upscaling software on M43 images produces results that approach native medium format resolution. The software can sometimes introduce minor visual artifacts. The M43 system maintains an advantage in portability, as a full medium format kit would be larger and heavier. He observes that elements such as composition, focus placement, and aperture selection have a larger impact on the final image than the camera format.

You might be surprised how good AI upscaling is. At our booth in the Long Beach scuba show for Southern California Underwater Photo societies, we had 16 x 20 prints from all manners of camera ranging from Z9/R5 to TG6 to iPhone. Even the lowest native resolution devices could make nice prints.

Bill

Is AI upscaling actually your images though? I tend to believe it's not.

4 hours ago, Grantmac said:

Is AI upscaling actually your images though? I tend to believe it's not.

Why would believe it's not your image? This is not the AI that will make an image or video for you based on a description, rather it's a specialized software designed to reconstruct detail that was present but not recorded and it's trained pairs of low and high resolution images to help make the predictions. It basically works out how to draw lines between your existing pixels rather than using a straight line or fitted curve that is used in standard upscaling. It's true it predicts the small details from upscaling but the lighting and composition is still your image, I tend to think of it as improving appearance of fine details that get lost in artifacts from standard methods of upscaling.

1 hour ago, Chris Ross said:

Why would believe it's not your image? This is not the AI that will make an image or video for you based on a description, rather it's a specialized software designed to reconstruct detail that was present but not recorded and it's trained pairs of low and high resolution images to help make the predictions. It basically works out how to draw lines between your existing pixels rather than using a straight line or fitted curve that is used in standard upscaling. It's true it predicts the small details from upscaling but the lighting and composition is still your image, I tend to think of it as improving appearance of fine details that get lost in artifacts from standard methods of upscaling.

The Ludd in me is escaping!!! How is it possible "to reconstruct detail that was present but not recorded?" Is the software cognizant of the world? I put forth that if you capture an image and then it is manipulated by a system you do not control, AI, to be something other than what was captured, it is adulterated and not the work of an original, therefore not your work.

Since photography went digital there has been software that aimed for similar with varying degrees of success and some of the software was marketed as intelligent or AI. Before photography went digital, folks did it the slow way with paintbrushes, masks, scissors, magnifying glasses and a lot of skill.

Based on its training and the data available a generative AI makes a well informed prediction of what the missing detail should be. With good training such a prediction can be accurate. There will be cases where the prediction is plausible but wrong, and further cases where the prediction is stupid.

All generative AI image 'improvement' has done is to take another step forward in a process with a very long history.

I mainly focus on underwater videography but sometimes I could use some good quality still images. Couple of months ago I saw a video (see below) about using video files to make high resolution still images. I'm quite convinced:

Anything which predicts to create an image isn't the original image.

I don't personally want to take credit for an image which AI produced.

14 hours ago, Cromagnon said:

The Ludd in me is escaping!!! How is it possible "to reconstruct detail that was present but not recorded?" Is the software cognizant of the world? I put forth that if you capture an image and then it is manipulated by a system you do not control, AI, to be something other than what was captured, it is adulterated and not the work of an original, therefore not your work.

Well we better not use digital imaging then, a Raw image before de-bayering is not like what you see on the screen, it constructed by the Raw converter according to the secret sauce each camera manufacturer makes which uses interpolation techniques to predict what each pixel should be. Sharpening is also out as it changes contrast around edges adding or changing brighter and darker pixels to give the appearance of a more defined edge to details in the image. Noise reduction - we are predicting what is image and what is noise.

This is the level of changes that are being made in the AI re-sizing we are talking about. when we do re-sizing of an image it is also predicting what pixels lie between the known pixels. The standard methods that we have all been using in photoshop or Lightroom use interpolation methods to fit either linearly or a curve between pixels. It does this whether you are going up or down in size. This is just a different method to accomplish this task and the AI is manly about recognizing noise and other artifacts and not magnifying them and only working with the actual subject data.

The problem we have is that we are shooting subjects where there are no straight lines and representing them with lots of little perfect squares called pixels. This brings all sorts of problems like interference patterns, moire and avoiding jagged edges that should be smooth and throw into the mix noise which we need to separate from from the image data. Like it or not this involves computations and predictions that are used to convert all the ones and zeroes into something that is aesthetically pleasing. All of these tools are about dealing with limitations of the sensor and various artifacts that the technology creates.

I don't hold with this idea of the purity of a straight out of the camera image, this is just Canon or Nikon's interpretation of what processing should be done to the image rather than my interpretation of the image and the processing needed to achieve this. Film is really no different - it's just the film manufacturers secret sauce applied in an analogue situation rather than digital. I'm not talking about cloning and adding things to images - just enhancing what the camera has recorded and dealing with all the noise and other issues in the data.

I'm a complete luddite when it comes to the AI that's booming around the place right now ( quite likely a big bubble getting ready to burst) and I don't use any of it. The task specific AI like this is a different story though, it has a definite purpose and the business model is relatively sound with development paid for by licensing fees.