So, in the interests of a diary of the Sabriel short project, here’s where I’m at (not very far), tagged for convenience or Luddite principles or whatever:
So I’ve already addressed the question of why I’m playing with this, kind of, but really I’d been thinking on and off for probably about eight months that I’d love a Sabriel movie, and we’re unlikely to get one and if we did it would probably suck because fantasy film making at the moment has chronic marvel industries irony disease. And I pondered making, say, a trailer with blender 3D (this is a 3D design program which I use for other stuff, very useful) but if I make big complicated files in that my computer starts making scary noises. Then AI video got good enough to be interesting so I am at least going to have a go at making a few scenes. Onwards!
I have a fairly particular internal eye that I want to satisfy and this is surprisingly difficult with AI. Possibly because I’m terrible at prompting, but more, I suspect, because I know exactly what I want and have to nudge programs away from making their own assumptions because either I wasn’t specific enough or I was so specific in one go that I broke the whole thing. A large part of this seems to be that AIs are, as we all know, Large Language Models, which means they neither think nor (this is important) see. Hallucinations and image drift seem to come about in part because of a loss of data that happens when you import an image, the AI translates it into backend text, and then translates that back into an image —just consider how many thousands of black bobs (hair) there are on the internet and you get the problem with letting a program decide what that means with a degradation in consistency at every step. So in order to nail consistency in imagery, what I’m doing at the moment is making a lookbook of consistent references that I can feed back into my prompts to generate images that will act as a sort of storyboard and that can then be animated into a scene.
But first, let me give you part one of a history of fucking around.
So, character design. Obviously I started with Sabriel, because she’s our protagonist, and I plugged a description of her that was basically cut from the book into an AI. The result was, bluntly, terrifying:
Could this be an Abhorsen? Ye-es. Is it Sabriel, a sweet, competent private school girl who doesn’t know how to talk to boys and is not emotionally ready to step into the boots of being a necromantic duke? No.
(This was all months ago and I’ve since lost the exact prompts, I’m just giving a history.)
Attempt two, which tried to at least make her look a bit more age appropriate, wasn’t as bad, but while Sabriel should look a bit unusual due to the death-bleached skin, she’s not even close to the scariest thing in the books.
Version three was still not great, because ‘really, really pale but still, like, cute,’ is a fine line to walk, and her costume was wrong, but she looked more like a protagonist that wasn’t incredibly villain coded, so I spent some time tooling around with her to see if I could get something resembling charter marks. This didn’t work, whether in the form of various prompts for a magical mark, taking the generated image and editing it myself and prompting that this was either a tattoo or makeup that had to be consistently incorporated, or anything else that I can recall, so I resigned myself to the notion that things like baptismal marks would have to be added in manually, which was still much more feasible than doing the whole thing myself.
She also ended up, entirely by accident as a result of something like ‘1920s bob’ in the prompt, with finger waves! Which is incredibly cute and not at all workable for someone who is either in a school uniform or filthy armor up to her knees in corpses. But she probably wears them when she’s attending garden parties in Corvere. Behold CW Sabriel.
Good enough to be going on with for the moment, said I, and moved on to being annoyed at how elements kept melting into each other and wondering if an animated style would be more forgiving of the mid and post processing that I’d have to do eventually. So I prompted this character ‘anime style,’ accidentally hit the video button, and ended up NOT with a line drawing version of the above, but with a grinning, skipping monstrosity that was unintentionally hilarious.
Whoops.
Okay, I think, maybe I need someone for her to interact with a bit. Enter CW Touchstone, in more shiny and less shiny v1 outfits. By the WAY, GARTH, do you know how much an outfit described as a kilt wants to autocomplete as ‘with a sporran!’ something you explicitly didn’t include???
This dude, for the record, is probably going to be supplanted with an imo cuter v2 Touchstone, mostly because (despite him being described as not exactly handsome) my personal headcanon departs from this in favour of the opinion that being the illegitimate son of a royal favourite probably means his father was pretty, and a teenage girl power fantasy means a pretty love interest.
Test shot of two teenage dorks pretending that they’re not not getting along with each other on their post apocalyptic camping trip. Looks like shit like a low budget made for tv series. AI seems to default to a stage blocking layout of characters always facing towards the camera, which is understandable but boring to watch. On review of the video, which is sort of meaningless vague interactions, I actually did manage to get the AI to keep the baptismal marks in place, but only by making them black, which I really disliked. (‘I’m just gonna have to manually edit this shit.’) Plus a bonus still of CW Touchstone reacting in a completely uncharacteristic way to being shown the Book of the Dead while not being a necromancer.






