Maximum megapixels (resolution) of ai generated images

[Devnish]

Adobe stock generative ai guide pdf stated that "Use the Enhance tool in Lightroom to increase the original file size, but don’t go beyond 6 to 8 megapixels".

But when I read some questions in the community, seems like people are uploading files well over this limit. Sometimes I reduce the image quality to meet the maximum of 8 megapixels when export in the Lightroom.

My question is can I upload upto 100 megapixels and 45 MB file size if there are no artifact in the ai generated image. Actually don't want to go file size that big. At least can I upload lightroom enhanced image around 12 megapixels?

[daniellei4510]

It appears you are referencing Lightroom's Enhance tool and ONLY Lightroom's Enhance tool as opposed to various available AI upscalers. I could be wrong, but I don't believe the Enhance tool in Lightroom is using AI to enlarge images, but is more likely based on Photoshop's Preserve Details 2.0. So it may be recommending that you limit upscales to a maximum of 8MP due to inherent shortcomings compared to AI upscalers. If you can beyond that without introducing artifacts (which is doubtful but I suppose possbile), then sure, you can probably submit 12MP assets.

[Abambo]

I could be wrong, but I don't believe the Enhance tool in Lightroom is using AI to enlarge images, but is more likely based on Photoshop's Preserve Details 2.0.

By @daniellei4510

I doubt that any upscaler uses AI for upscaling. My guess is that they use sophisticated algorithms that may have been developed using AI as a helper. We used such tools to optimize processes, called the resulting tool xyAI and went to market. It's marketing. 20 years ago we would have called it fuzzy logic. 

The Lightroom upscaler works specifically well with raw files, if I remember well,, as their format anyhow needs heavy processing to get a nice colour image. 

[daniellei4510]

Agreed. When Gigapixel AI updates, it installs new "reference models." That is, it doesn't scour the internet but uses internal reference models to assist with the upscaling. So it's...semi-AI? In any case, it seems to work...most of the time. For example, Gigapixel AI has a tendency to oversharpen, especially irises and strands of hair. Fortunately, Gigapixel AI creates a new layer of the sharpened image while leaving the original (and of course softer image) untouched. So masking out the hair and irises is often required. I also keep an entire library of AI-created irises that needed little if any editing and use them as needed when the eyes aren't perfect after upscaling. Blue eyes, brown eyes, green eyes...the whole gamut.

[Abambo]

Oh, it may well treat landscape images different from portraits different from buildings. Upsizing an asset makes it blurry. The art is to correct that, by keeping hard edges (high frequency) hard. 

[Nancy OShea · Accepted solution]

Upscaling from low-res to high-res can introduce artifacts and other unwanted side-effects.  Aggressive upscaling can make matters worse.  Whenever possible, start with a high quality image that contains at least 4 million pixels (height x width).

Adobe Stock Photo & Illustration Requirements:

• Image format: JPEG 
• Color space: sRGB
• Minimum image resolution: 4MP (megapixels)
  
• Maximum image resolution: 100MP (megapixels)
  
• Maximum file size: 45MB (megabytes)

https://helpx.adobe.com/stock/contributor/help/photography-illustrations.html

IMO, Gigapxel AI is the best image upscaler for AI.  Whether you use Gigapixel or Lightroom is up to you. But don't go too heavy with upscaling. Afterwards, carefully inspect every quadrant at 100-300% magnification. Correct all defects if you can.

https://www.topazlabs.com/gigapixel-upscale-ai-generated-art

Hope that helps.

[daniellei4510]

Totally in agreement when it comes to Gigapixel AI. Their latest version, in fact (7.0.4), has a specific recommended setting for AI generated images. 

[Abambo]

You need to know that any recommendations (!) are done for the time when they have been issued. You have a fast evolving technology. If you think your assets are good enough at a higher resolution, go for it. But if they aren't, do not submit, it is not an advantage to get assets accepted, when they do not meet the requirements. The first buyer will complain, a lot of comlaints will get your account closed.

I doubt that you can produce a generative AI at 100Mp, that does not expose artefacts introduced by upscaling. But if you can generate high resolution assets, you should even nit need to upscale to meet the minimum resolution of 4Mp.

Please note that upscaling does not make your assets better. It amplifies any defect that may be present.