Book Cover Testing with AI

[Richard]

I found a free AI tool, that tests fiction book covers.
https://joelbooks.com/ai-fiction-cover

[T.A.]

Wow this is just like an AI SEO Audit that tests per parameters. so cool thnx!

[Harald, The Swimmer]

Thanks for pointing to this, Richard!
While it’s an intriguing idea (especially since I’m investigating AI Voicing for audio), I have some concerns...

First, I tried it with four of my fiction covers. Result: 3 “Lows” and 1 “High”. Curious, I looked into their methodology more (explained on the same link). And I see the problem.

Their “pre-trained models” (determined by one book marketer and two cover designers) do not reflect low or high cover quality well enough, in my view. While the Low examples are uniformly pretty bad, the High ones are all over the place, from very good to very bad (in my opinion). Just go to the link and see if you don’t agree with me. I could easily put about half of the “good” ones in the “bad” group in light of today’s online marketing needs. I get the sense the 3 judges here are either Old School or just plain Old! ;-)

So while it’s an interesting exercise, my opinion (currently) is that:
1. It’s too simplistic. Only two choices: Low or High. They need more factors to reveal, e.g., Brightness Contrast, Title Readability, Colors, etc.
2. The dataset is flawed. They either need more samples or more judges or both.

Maybe it will get better over time. Thoughts???

[T.A.]

While I haven't tested any covers, I comprehend the AI would have those weaknesses but only because this application of AI is new I'm guessing.

I've been noting lately loads of covers (fantasy romance genre) that are very much like home-craft Photoshopping bcz I do photoshop and can see the indications. I suspect an AI this early in the game would assess for Color and Composition first.

[Judy]

Agree with all the concerns mention, but it is still very early days for AI and book covers.

[Paul]

And, as the limit with so many "learning" algorithms: the algorithm will tend to favor similar results. Any system using statistical "intelligence," artificial or not, tends toward homogeneity. The modern equivalent of groupthink.

The reality when it comes to covers is that there are some rules which shouldn't be broken, some rules that can be bent if you are purposeful… and sometimes the outlier works for reasons that make no sense.

But this is an interesting read. Thanks for posting, Richard!