How to increase the resolution of AI images to an acceptable level for stock submissions

[artycrafter]

Hi, I've been a successful photography stock contributor for many years creating historical costume images predominantly used by authors on their book covers. My shoots involve a lot of work and expense and so I thought I'd play with AI to generate a wider variety of images using my existing photos as the reference images. While it's been somewhat hit and miss I've produced some interesting images that I think will appeal to my customer base. 
As I'm used to a 45mp camera and the resulting ultra sharp, high resolution images, I was disappointed with the resolution and sharpness of the AI images. I'm experienced with Photoshop so used the generative resolution upgrade, increasing both  the pixel dimensions and dpi, and the sharpening tools in an attempt to improve the resolution and sharpness. Thr results were ok, but not great.  The AI engines I used were Adobe partners nano banana 3 and GPT through the Firefly portal as these two provided the results I was visually most happy  with. 
I decided to submit 4 images to see what happened and they were rejected for quality reasons. As no specifics are given it's difficult to know exactly why they were rejected, but I assume it was resolution and sharpness. 
I've included a couple of images with this post for illustration purposes. I am not asking for a critique as I know the resolution and sharpness are not good enough, I am asking if it is possible to generate any image that has decent resolution and if it is, which AI engine should I use? I'd rather generate decent resolution images in the first place rather than trying to 'save' a crappy image in Photoshop afterwards.
Any practical advice anyone has that is based on their experiences would be most appreciated. 
TIA

[daniellei4510]

All text to image AI models that I'm aware of generate results ranging in the vicinity from 19 to 20 in. on the longest side at 72 ppi. I use Gigapixel AI to upscale my images and have had assets accepted that were as large as 30 inches on the longest side at 300 ppi. (I've since dropped to 20 or 24 inches to minimize subsequent editing).

That said, if I look at your top image at 200%, the people featured on the tapestry are poorly rendered, which would more likely be the reason for refusal. On the second image, the pattern on the puffy shoulder sleeves don't match as well as they probably should, and these are the things moderators appear of late to be taking a closer look at when it comes to AI submissions. So I don't think a lack of resolution was necessarily the reason for these assets to be rejected. But yes, Gigapixel AI is what you'll want to consider looking into when it comes to increasing resolution and retaining (and even improving) sharpness.

[Jill_C]

In the first image, the girl has no fingernails, and the weaving of the golden sash down the front of her gown is inconsistent. The image overall appears to be quite noisy, which might have been intentional on your part, but the moderators probably just see it as noise.

[Nancy OShea]

Resolution is just one small part of the equation. 

No AI engine is perfect. They all make mistakes, some worse than others. The more complex an AI artwork is, the greater possibilities for errors.  

Every quadrant of output should be closely examined for accuracy, logical rendering of details & angles, and removal of unwanted artifacts.  When artwork looks synthetic, it's often due to subtle inconsistencies that only humans can detect.

[Nancy OShea]

Image #1, pearls look slightly off to me.  Also, the bodice creases look odd. For comparison, look up Medieval corsets to get a sense of where the boning structure should be. 

[Abambo]

As you are mentioning increasing the dpi: First you increase the ppi value, second this is the most unnecessary parameter for stock photography. There is only one reason, why this parameter exists: to translate the pixel values into inches. So a picture with 2000points may be interpreted as a 100ppi image, meaning that it will be interpreted as a 20 inch length, or a 200ppi image, meaning that it will be interpreted as a 10 inch length. Nothing, and I repeat nothing has been changed in any quality related parameters. So for convinience, leave the ppi value at 200 or 300, whatever your prefered value is, and don't consider it anymore. It's completely irrelevant.

The resolution, however, is probably the most intersting parameter that you can touch up. But increasing the resolution does not create new information, but that information does need to come from the neighboring pixels and gets interpolated. There has been big progress in the course of the last years in this, increasing the resolution of assets intelligently, trying to keep the sharpness, without introducing new artefacts. But there are limits to this method, as you have experienced.

For some uses, those limits are not important, for some other applications the limits and the loss of quality is a major issue. As for this reason, do not increase the size too much, so that the asset, when checked at 100% still holds the necessary sharpness, without introducing artefacts. But this means that your source imge needs to be perfect, which is rarely the case with generative AI. You need to fix rendering errors first, then scale up and fix upscaling errors and those rendering errors that you did not see during your first checking. It's not al all an easy and fast task.