Author Tools - AI cover art - state of the tech

 Cover art is one of the biggest barriers to entry in the indie author space. It runs from cheap to unobtainably expensive, from terrible to transcendent. Everyone has different tastes and visions, and opinions on style run the gamut.

But everyone agrees on one thing - it matters and it has to be good.

One of my biggest challenges and points of growth has been to understand and accept that covers that get a response on the market may not be what I prefer to see. At least in the affordable space.

Thus, when I saw indications that art AIs had ‘leveled up’ since the last time I played with one, I deemed it time to run some experiments.

I’ll start with a summary of my personal preference. I like illustrated art that depicts a scene from the book, or something close to one. The prime example of this is Michael Whelan’s cover for Sanderson’s Way of Kings. This is a book I bought because of the cover, when the blurb may not have been enough to convince me and I did not yet know the author.

Obviously this quality of art is out of reach for a beginner author, and I am not the only indie who has attempted to access the illustrated style by getting the best I could afford.

I spent the third most amount on this illustration of all my books, and it’s the best of the bunch overall, so far. It also had the greatest reception of all my illustrated covers. Nevertheless you can see (or at least I can see) a few issues. The pyramid ship is incorrectly scaled, and the Krashmobile is far too small compared to its description in book. Still, the picture gets the point across, and the colors are engaging. Could I have fixed my issues with it, or added more characters to the scene? Sure, but I couldn’t increase the budget.

General wisdom is, if you can’t afford an illustrated cover where one would be appropriate, get a photo real composite. Some indies approximate this by using 3D renders. In general, I dislike both and they do not draw me in.

While these are decent quality examples of both styles, something about them just turns me off. Would not click. That said, someone would, and people do. Thus, authors like myself must figure out how to meet our prospective readers where they are. For me, this has meant practicing my design skills and deciding how to source affordable images.

Enter AI.

Roughly two years ago I played some with Artbreeder, before they locked it down and started pushing their pricing tiers a little harder. In general, it did not produce art I found suitable for covers. But I did get a small set of abstract sci-fi pieces that I was willing to use if it came to it.

I’d say these are somewhere between old school and obviously indie, and frankly don’t leave too much to be desired other than giving off a more chill vibe than the books actually possess. So far as I saw, this is the best Artbreeder could do for our purposes. A pattern begins to emerge, however - the colors and overall vibe are spectacular.

Having left AI alone for some time, I was pleased to see recent posts by people using things like Midjourney and Dall-E 2. I also came across DiscoDiffusion, which can be accessed for free but takes some fiddling.

Without access to Dall-E 2, I’ve only been able to test the others.

Here are some tests with Disco Diffusion. Two are attempts at a photo real but stylized Bloodbane cover, with Castlevania art as an init. There’s one where I tried to get it to isolate a cyberpunk building for game art purposes, and then a general cyborg concept.

You can see the big weakness here is humanoids. Proportions, faces, angles. Everything is off. You can get the AI to do decent portraits, but generally we won’t find those useful for cover art.

Disco just couldn’t figure out how to portray my lumberjack warrior, nor could it really grasp where the axe should be incorporated. That’s to say nothing of trying to get it to render him using a polesaw as a weapon. On that the AI was just clueless, since there is no precedent for that in its trained images.

Once again, the color palette and vibes are beyond excellent. If the covers didn’t need to be coherent, these would be very close to serving their purpose.

Which brings us to Midjourney, which you may have seen me playing with on Twitter.

For reference, here are my two illustrated Hero’s Metal covers, which I love very much for what they are.

But I have to be honest - they do not fit in with what else is on the market. Would I click them? Absolutely. But not enough other people are. In the absence of an $800+ illustration, we need something photo real.

Now we’re talking. Abstractness remains, but the machine’s ability to generate realism while also stylizing has come a long way. You can see the emblem style is the most focused, while the landscape is the most beautiful.

As usual, colors and vibe are the best elements. But the AI’s sense of lighting has greatly improved, we made use of style matching, and the sense of photo realism has greatly improved. Book 4 in particular is still in iteration, but I wanted to include it as an example of what we can do with multiple generations and compositing, for posterity.

Here are two more pieces I generated, for stories I concepted in tandem with the prompts:

I consider this level of art more than usable, and in fact have updated Hero’s Metal 1 and 2 on Amazon. We will see if they generate more clicks/buys. In any case, the reception on Twitter was great, so I’m taking that as a good sign.

I’ve signed up for the Dall-E 2 access, and will report if and when I can. I will also follow up with tips for prompting such AI, should there prove sufficient interest.

The verdict then, for now, is that emblems, landscapes, and somewhat abstract pieces are generate-able and usable. Be aware that it takes some playing with the AI, which means subscribing to one, which means paying out. But I’ve got at least 4 covers here for a much more than reasonable price.

While we would all much prefer to pay talented, working artists in our circles for amazing and beautiful cover illustrations, I think this route for our cover art is not something we can easily ignore.

Questions? Critique? Hit me up!

And click on those beautiful covers above to buy Hero’s Metal and all my other fantasy adventures.

No comments:

Post a Comment

The Mythos Project, or Conservative Retellings

  With the rise of big tech censorship and open calls to silence voices that do not agree with the zeitgeist, I’ve been slowly coming to a p...