AI-dventures in a Strange Realm

First of all, apologies for the terrible AI pun in the title: it’s just a shameless continuation of an established tradition. 🙂

Earlier I posted extensively about my work at Apocalypse Studios on illustrating a fan-favorite lore video and podcast series called Chronicles of the Grand Inquisitor, using a continuously improving proprietary workflow assisted by an array of generative AI tools and techniques.

Despite the raging controversy surrounding generative AI I have been a vocal advocate for the technology, and I believe in the hands of skilled creators it unlocks of wealth of new creative possibilities.
I have been trying to demonstrate this potential through my own work, all while having an unprecedented amount of fun in the process.
Regardless of what the fire-and-brimstone outrage mob wants you to believe, I found that there are 3 basic axioms to so-called “AI art”:

• AI without an artist is an instrument without a musician.
• AI and its training data are two different things.
• Relying on prompts will not get you very far.

With that out of the way, let’s talk about the proverbial “cool stuff”.
In the latest episode of the series, titled ‘The Ark,’ the main characters, Alaric the vampire and Amarax the lich, find themselves in the Etheric Sea. This strange realm-between-realms is a steely blueish-gray domain filled with enormous, perpetual storms and violent currents of magical energies.
They travel on an Ark ship constructed out of parts of a dragon’s skeleton, and they visit the monolithic, octahedral black temple of Anu Maht, which conceals gruesome secrets.

In other words, a lot of highly specific stuff with a very particular visual language that basic prompting can’t even begin to touch (of course anyone is welcome to try… 😉 )

Consequently, the workflow I used for this particular episode was markedly different from all previous ones:

• I created hand-made reference images for color keys, textures like the cloud cover or the black stonework of the temple, character likenesses and facial expressions, costumes, and more (I’m planning to do a separate, more in-depth post about these soon).
• To get the layout and composition of the shots right, I used a combination of simple storyboard sketches and placeholder images generated by high prompt-adherence models like Adobe’s Firefly and Leonardo’s Phoenix.
• These shots looked like a messy scrapbook of rough speed-painting snippets and bland stock photos, but they had the poses, objects, and general placement of elements dialed in.
• Relying on these layout shots as mere guidance “proxies” I used a ControlNet to style-transfer the hand-made reference images to the scaffolding structure provided by the layout shots.
  One of AI’s game-changing features is its ability to separate, extract, and recombine an image’s expression (composition) and its appearance.
• Some shots, like the Ark ship, required more elaborate handcrafted reference images and multiple iterative passes to get right.
• I generated a number of variants for each shot, influenced by a custom LoRa model to reintroduce some of the stylized pseudo-cartoon look from previous episodes.
• I tweaked and assembled the final shots in PhotoShop, some of which are composites of up to a dozen variants.
• After a bit of final manual touch-up and color correction the shot was ready to go.

I’m obviously biased, but I think the result is a series of images with remarkable stylistic consistency, and, despite the very specific and challenging subject matter, an impressive level of adherence to the written material.

But why go to these lengths and use these esoteric processes to create these images instead of just making them by hand, you ask?
As I’ve voiced many times, I believe generative AI is the first tool to enhance the temporal aspect of the creative process rather than the mechanical one. This is very much the case here too.

Instead of investing weeks, if not months, to fully handcraft 30 shots at this quality and level of detail – which, given the cost and the ROI involved, would be a non-starter – the entire set was created in a matter of 4-5 days using this workflow, which is a value proposition that’s hard to ignore…

The images:

Here’s the full episode, with the great narrative and excellent voice acting:

Also, check out the follow-up / second part to this post with even more images.
…and finally, a little bonus, a goofy twist on the classic yelling-at-cat meme for the social media promo material of this post: