Click or swipe to change slides · Open full screen ↗

What Developing 20+ Open-Source Tools Taught Me

At the Sydney Nuke User Group in November 2026, I talked about something I'd wanted to do for most of my twenty years in VFX: build my own tools.

Tutorials, programming books and maths courses never quite got me there. AI did, although not at first. My first attempt in September 2025 failed badly. Five months later, the models had improved enough to make it work.

Since then, I've built 27 open-source Nuke tools, ranging from small workflow fixes to native renderers. Each project taught me something that fed into the next. I've been documenting those lessons in NukeDevRules, so every new project can build on what the previous one taught me rather than starting from scratch.

The technical deep dive was InstanceRender, the renderer behind the forest scene in the talk.

That forest consists of 48,000 instances of just 33 meshes. Flatten those instances, as ScanlineRender2 would, and you're looking at nearly seven billion triangles.

InstanceRender takes a different approach. It stores each mesh once, bringing the actual geometry down to around eight million triangles. The result is a frame rendered in 33 seconds on a single gaming GPU.

CopyToPoints scatters the instances as standard USD, while InstanceRender ray-traces them on either CPU or GPU. It supports full AOVs, light splits, per-instance Cryptomatte, deep output and motion blur.

From the first commit to release, the project took just five weeks, supporting six different versions of Nuke.

I also spent some time talking about where AI gets it wrong.

At one point, my GPU build shipped with no shadows at all. The tests still passed, because the test itself was incapable of failing. It was a good lesson in the difference between something that looks correct and something that actually works.

I think of AI as a super-smart assistant with an enormous amount of confidence. It can produce what most good engineers would write, often remarkably quickly. But when it reaches territory where there isn't much prior work to draw from, it starts guessing.

It can't feel when a knob layout is wrong, when an interaction doesn't make sense, or when something simply doesn't feel right. An artist can.

The real skill is learning to brief AI in the language of the discipline, get it to check its own work, and, most importantly, stay involved in the process. The human still needs to understand what is being built and why.

Every tool I showed is free and open source, available on GitHub and Nukepedia.

As Phil Tippett discovered on Jurassic Park, the tools may change, but the artist's eye doesn't.

Download the deep dive test scene.