Mechanical Labyrinth

This is a project for myself with a few goals: improve general proficiency with Blender, with a soft focus on hard surface modelling; tell a story – or rather, establish a world and atmosphere – through setting with static camera shots and environmental animation only; and produce a finished video of around five minutes. Essentially, learn stuff, make stuff, finish stuff.

Story outline

Background: A century ago – our not-too-distant future – an autonomous factory controlled by a poorly-implemented AI went out of control, growing and displacing all living things before its rapidly expanding frontier. Human pushback was ineffective and it overgrew the whole contiguous landmass of the continent.

Present Day: Humans have given up the “Mechanical Labyrinth” as a lost cause, and entry is forbidden. The machines, expending great effort and producing nothing in particular, have hollowed out the ground beneath for tens of miles and seem poised to eventually undercut the whole crust, but for whatever reason growth has slowed (perhaps the AI has overreached itself, or is dumping energy into a futile effort to expand into the mantle).

The labyrinth is now in decay, and we see it in its dying years. It is in one sense a hopeful thing, the existential threat being averted, while in another it is bleak, as the rusting husk forming a scar over the Earth will be a lasting testament to a great folly.

Basically we’re exploring a sort of brutalist Aperture. The general concept came first, and the above story is there mainly to motivate in fleshing out details. (and while it reflects some anxiety I have about future technology, it is only a basis for a setting and not intended as any sort of moral tale.)


The video will be a series of disjoint scenes from within the labyrinth, each about 5-10 seconds long. The nature of the setting provides room to try out a lot of different little ideas that wouldn’t necessarily stand alone. Scenes will be accompanied by ambient soundscapes that vary distinctly from shot to shot.

This sketchbook piece gives an example of the kind of little vignette I’ll be making.

All assets will be my own work except where specified.

I expect to have somewhere around 40 scenes/shots. Production time I expect to vary from a week for the lightest shots to a month for the heaviest. I hope to take no longer than a year to finish, but things happen.

“weird train”

I started on this after watching some of Josh Gambrell’s tutorials, and followed the same general approach. Not using any paid modelling addons for now though.

This train engine took a long time to do, in large part because I unwrapped it and manually painted curvature maps. I couldn’t find a good way to automate it, but I think I should look again at ways to turn object-space normals into curvature. In the meantime I’ll probably stick to AO or pointiness-based approximations for the sake of time.

I had a little success with textools. It expects hi- and lo-poly inputs and I’ve only got one object (sort of mid-poly), so I made a hi-poly using extra bevels. Didn’t end up using the result but I could revisit that too. Also tried Armorpaint; the result is slightly different each time you apply the mask which sketched me out.

Tip: texture painting on big textures is very slow. Blender crawled painting the main body (4096px2). For all the miscellaneous parts I used a UDIM image with lots of tiles no bigger than 2048px2. Pleasantly surprised because UDIMs are new and I figured they’d be janky and delete stuff when you’re not looking, but they’re fine. Only issues are that it seemed to require more delicate handling when resizing the first tile (can’t quite remember the what/why), and it’d be nice if you could save only the changed tile rather than the whole lot (which I assume it does because all the thumbnails in the file browser get remade each time).

Tip: texture paint in orthographic mode. In perspective mode, particularly with flood fill, you can miss spots or create big random lines and spikes all over the texture which can go unnoticed. This might be affected by topology, but you can try it on a default cube with the camera zoomed far out.

Material isn’t finished yet but it’s coming along. One thing I want is to limit materials and imperfections to stuff that makes sense in a humanless environment, so simple bare metal with accumulated grime, dust, rust, wear and tear, but no paint jobs, fancy flourishes, handprints, graffiti, etc.