Nikon Rumors - comparing Gigapixel AI and Photoshop

[Tibor Hrnko]

Interesting article on the subject of post processing digital photos.
Comparing Topaz Gigapixel AI vs. Adobe Photoshop

Executive summary:

• Gigapixel AI creates random artefacts in resulting photos
• Default values used by Gigapixel AI are not optimal
• Usage of Solarize filter is recommended to spot such problems
  • It is a generic tool to evaluate smoothness of any picture, even before PP is done
  • Method to create such filter is described in the article
• Examples of using Photoshop and GAI are presented in the article.

 

[bleirer]

Looks like it specifically is comparing upsampling using preserve details 2.0. Be interesting to put the Photoshop super resolution in the mix.

 

[Tibor Hrnko]

Looks like it specifically is comparing upsampling using preserve details 2.0. Be interesting to put the Photoshop super resolution in the mix.

I was missing that one too. But I like the method used more than the actual results.

I have found so far neither GAI nor Photoshop super resolution doing something special to the original photo. I am using PSR from time to time on a picture to enhance level of details, but it never was providing exceptional results specific to its use of Machine learning. I am using high pass filter much more than PSR to enhance small details at the end of PP flow.

 

[bleirer]

I was missing that one too. But I like the method used more than the actual results.

I have found so far neither GAI nor Photoshop super resolution doing something special to the original photo. I am using PSR from time to time on a picture to enhance level of details, but it never was providing exceptional results specific to its use of Machine learning. I am using high pass filter much more than PSR to enhance small details at the end of PP flow.

I think it's apples and oranges. The upsampling is for when you need more pixels and the programs use the surrounding pixels to extrapolate. You never get more detail than the original had. Sharpening is a separate thing. One could still use high pass or whatever on a upsampled image. You can never get more detail by sharpening, just more edge contrast.

By the way, in my experience smart sharpen using the 'lens blur' mode is very like high pass in that it detects edges, while the Gaussian blur dropdown is more like unsharp mask.

 

[Tibor Hrnko]

...You never get more detail than the original had...

My understanding of AI interpolation is the one using qualified assumptions of underling pixel logic of a typical picture details
My understanding of bicubic or any other classic interpolation is using the surrounding pixels to generate new ones.

• In that respect AI should/could produce more details if properly trained. But I have not seen that so far in any of the 'AI' solutions available
• I was unprecise in my English stating that high pass enhances the original image. Apologies for my language skills and thank you for commenting.
• the kind of artifacts highlighted by the author of the Nikon Rumors article is problematic even as a tool to 'visually enhance' the PP
• The main takeaway for me is the usage of solarization method to get estimate of smoothness of picture gradients

 

[bleirer]

My understanding of AI interpolation is the one using qualified assumptions of underling pixel logic of a typical picture details
My understanding of bicubic or any other classic interpolation is using the surrounding pixels to generate new ones.

• In that respect AI should/could produce more details if properly trained. But I have not seen that so far in any of the 'AI' solutions available
• I was unprecise in my English stating that high pass enhances the original image. Apologies for my language skills and thank you for commenting.
• the kind of artifacts highlighted by the author of the Nikon Rumors article is problematic even as a tool to 'visually enhance' the PP
• The main takeaway for me is the usage of solarization method to get estimate of smoothness of picture gradients

That solarization was cool. Looks like high points in the histogram made low by the curve and vice versa.