Quote AI Jobs

Flame Between Renders
1

as AI requests are starting to flood us completely, and actually ideas are comming from directors based on capabilities they have seen in some AI demo…

I dont want to go into the depth of why this all sucks or doesnt , but we cant be standing still and need to “play ball” with our clients.

My current approach is to quote every AI request in 3 Ways

  1. We find a commercial tool (there are updates/new tools EVERY WEEK) that does what we want , this is cheap but we cant even start to predict the outcome, so we cant GUARANTEE this will work and we might be left with garbage and still waste money on it, its “worth a shot” . Adding some comp to combine different AI tool results e.t.c

  2. We train our own models, possibly do some combining of traditional and AI tools to get somewhere, mid level approach, higher chance of working if 1) doesnt and can use stuff from 1) to enhance it but its still parts blackbox AI that we cant fully controll, so lots of trial and error

  3. sure-shot this will work 100% , we do it fully classic we can controll everything but itll be CRAZY expensive , lets say instead of a deepfake like 1 and 2) this is fully digi-double hand animated.

How do you tackle the pricing and stuff for these kinds of jobs? its so difficult not knowing what could happen and the clients mostly dont want the risk …

2

Seems like the right way to think about. Especially given expectations and state of affairs.

I would just give them two options:

A) Use best AI tools available, with additional comp/cleanup to make results usable. Include some R&D time, as it’s a fast moving landscape. Caveat that this is best effort, and no guarantee on quality outcome.

B) Classic (non-AI) approach.

That way they see both options in contrast and can choose. If it works, good for them, if it fails, make sure the understand the final price is A+B. There is a cost to taking a risk.

You can allow them to manage risk, by offering to go down path A to certain checkpoints. If it doesn’t look promising, they can abort and switch to B before they waste too much money.

And lastly, deconstruct the job and see what part AI is used for. AI has a higher chance of success if you use it just for certain elements (background plates, etc.). I think everyone wishes you just write a prompt and get a finished high quality product. The reality will be that you use non-AI for the critical components, but make up ground on all the mundane stuff.

If you think of the BMW video the guy made with RunwayML v3 - maybe the b-roll shots where the car is driving in the distance on open roads is AI generated, but the close-ups where details are more important is done in studio. And so are certain elements you can comp in. Any element you can skip expensive camera work on can give a bit of time to make up for some AI quirks and integration.

Are there any AI tools that give you elements with alpha channels? Or tools where you can control the camera path or get a camera track? Or tools that provide 3D geos? Probably not, but that would be super useful. Then you could take the ToysRUs film - and maybe build the store scene in a 3D app/CG render, comp in the old cars in proper perspective, and get an AI animation of Geoffrey, but film the kid as real actor and integrate. Finding a middle ground of efficiency, and quality. Just thinking out loud now…

1 Like