Netflix is doing a lot of AI stuff behind the scenes - everyone is, of course, but their first public foray into this world has been by releasing VOID, an inpainting model.

Inpainting is the process of removing unwanted parts of an image, such as a bit of a C-stand being visible in shot, or an unwanted bystander who has made it into the filmed footage.

It's an area of VFX where AI really shines, and we've built our own template using WAN to do it.

But to survive in a fast moving industry, you need to be able to adapt, and if a better option than our template presents itself, I'm more than willing to throw that away and switch to the better option.

And it's with that approach that we started using Netflix's VOID. Their unique claim is that VOID's paint-outs don't just remove the objects, it also removes their impact from the world. So that could be shadows or reflections, but really weirdly, also physical supports. So if you were to paint out the legs of a table, the surface would clatter to the floor. That's going to be a real headache each time I'm using fishing wire to suspend an object deliberately - remove the wire, and rather than being suspended by magic, the object just falls!

But that's not my issue with VOID. It's more that it's just...it's just not very good. There's nothing wrong with that - it's free and it's new, so if they keep pushing it, it will improve. Also inpainting often doesn't need to be that good, especially if you're removing something from shot to replace it with something else (e.g a stand-in actor, who will be replaced with a digital creature).

Most AI developments these days fall into the category of being not very good - or perhaps more accurately not very useful - but in order to stand out from the rest of the noise, they need to promise the earth, which results in a massive gulf between hype and effectiveness.

Influencers need to do the same, of course - you don't get many clicks for saying "here's a new update that's sort of useful."

Everything always comes back to the same thing: create shots, create films. As the need arises, try the tools that you think will help. When you have a real problem to solve, you're going to find out very quickly how useful the tool is.

For any tutorial, tool, or template we develop inhouse to release on the Shoot First platform, that's the one unbreakable rule. If you haven't tried the tool successfully on at least a couple of different shots, you don't have a tool yet. We're too small a team to have an RnD department, but even if I had 1000 artists, you'd have to smack me pretty hard to convince me to set one up, because I've seen the pattern too often. The second you step back from the front lines, your solutions are theoretical. Schrodinger's solutions - both useful and totally useless at the same time. You can't know until you try.

With that in mind, we've updated our inpainting template. One thing we were hitting was that WAN uses a metric tonne of memory (which, to VOID's credit, it doesn't). There's not that much we can do about that, but one big change is that we've implemented the same auto-cropping system that we use in the masking template. So if you're just painting out a fraction of the shot, then the template is going to become much more efficient.