What I built
A 40-second short film. A Jōmon dogū — the clay figure with snow-goggle eyes — says four sentences about what it was made for, its eyes light up, marigolds bloom at its feet, and the title card lands.
Every stage is local:
| Stage | Tool | Runs on |
|---|---|---|
| Script (126 Japanese characters, first person) |
Llama-3-ELYZA-JP-8B Q4_K_M via llama.cpp |
127.0.0.1 |
| Voice | GPT-SoVITS v2, fine-tuned on ~15 min of my own speech | local GPU |
| Image | Blender 5.2 LTS / EEVEE, 960 frames at 24 fps | local GPU |
| Edit | ffmpeg | local |
No hosted inference, no API keys, no cloud render. The narration was generated against a loopback llama.cpp server; the model file sits on the same disk as the render.
Why a dogū
The challenge asks for a ghost in the machine. Most entries will be a chatbot in a terminal. I wanted the Japanese answer instead: a dogū is a face made to be looked at — the goggles are a mask, the body is a vessel — and it spent a few thousand years underground before anyone read it back. The "ghost" is not new software. It is an object that was already a mask before it was a model.
So the film gives it exactly one thing it never had: a voice, and forty seconds to use it.
How it works
Three scripts, each re-runnable from scratch.
1. The script. A single prompt fixes the beats (I was a vessel for seeing → I slept → I am looking again → marigolds mark the ones who come back → a question at the end) and the length (120–180 characters). The model overshot or undershot on most seeds — 114, 128, 92, 80, 85, 78, 126 characters across seven runs — which is a useful reminder that "please write 150 characters" is a suggestion, not a constraint. I kept the one that landed in range and removed a parenthetical so it reads aloud.
2. The voice. GPT-SoVITS v2, trained earlier on a small personal dataset. For the dogū I shifted the output down two semitones and slowed it back to the original duration:
ffmpeg -i narr.wav -af "asetrate=32000*0.891,aresample=32000,atempo=0.858" narr_dogu.wav
The weights and reference audio stay offline — they are personal. docs/voice-setup.md records the procedure instead.
3. The image. This is where the actual work was.
Two things that cost me the evening
Millimetres are not metres
The dogū mesh comes from a 3D-print project, so it is modelled in millimetres: a 250-unit figure, a camera 660 units away, lights 400 units away. Blender renders as if one unit is one metre, so inverse-square falloff treats the key light as 400 m away. Ordinary wattages render pure black.
The fix is not "add more light until something appears". I measured a ladder:
| Light multiplier | Total energy | Result |
|---|---|---|
| ×1 | 11 kW | black |
| ×100 | 1.1 MW | nearly black |
| ×1000 | 11.1 MW | correct |
| ×20000 | 222 MW | blown white |
One LIGHT_SCALE = 1000 constant, documented with that table next to it, ends the guessing.
The arms were lying on the print bed
A print model arrives in print orientation: every part separate, each one placed so it needs no supports. The arms were 133 mm below their sockets, sitting on the base like dropped tools.
Rather than eyeballing it, I measured:
- fit a sphere to the arm's end vertices → ball centre
x = ±29.1, radius ≈ 7.9 - find the torso polygons whose normal points away from the torso centroid → the socket cavity, centred at
x = ±20.9, z = 81.4 - the arm's existing X was already "socket centre + ball radius", so the fix was a pure translation: +133 mm in Z
And then a trap worth writing down: after ob.location += delta, matrix_world and bound_box are stale until bpy.context.view_layer.update(). Without it, the pedestal, the flowers and the camera all get placed against the pre-assembly pose — the figure floats 100 mm above its base and nothing in the log complains.
Framing, numerically
Four renders went into guessing whether the marigolds were "in shot". They were in the scene, at the right height, correctly lit — and below the bottom edge of the frame. The thing that settled it was one projection dump:
from bpy_extras.object_utils import world_to_camera_view
pts = [world_to_camera_view(scn, cam, ob.matrix_world @ Vector(c)) for c in ob.bound_box]
# NDC y < 0 means below the bottom edge
If a prop "looks invisible", compute its NDC box before touching the material.
Why open weights matter here
The submission rule was "works with the lights and the internet off". A hosted model cannot satisfy that sentence, no matter how open its weights are. Keeping the whole chain local also made the piece cheaper to iterate: seven script generations and about twenty still renders cost nothing but time, so I could throw away the wrong ones — which is most of them.
The trade was quality on the writing: an 8B Japanese model wrote a usable 126 characters, but it would not obey a length limit, and it wanted parentheses. The interesting part of local-first work is not that it beats the cloud on quality. It is that it stays available, and it lets you redo things.
What I'd do next
- Eyes: the glow spheres had to be moved 2.5 mm forward, out of the socket cavity, or they render as dull holes at any energy.
- The window between "correct exposure" and "blown white" is about 1.5 stops at this scale; the final cut sits at exposure +0.2.
- The narration is 27 s in a 40 s cut, so the last 8 seconds are the title card — that is a design choice, but a longer script (150+ characters) would be better served by a 50 s cut.
Agent session
Built with Hermes Agent (marcos profile) driving Blender headless, llama.cpp and ffmpeg from one session; the still frames, the light ladder and the joint fit were all measured, not guessed.
Licence
Code and text MIT. Stills CC BY 4.0. Voice model and reference audio are not distributed.
Top comments (0)