Crew and equipment: who does what during the jam
Thu Nov 12, about 00:45 ET. Source of truth for machines and tools: studio-pipeline/registry/TOOLS.md (Machines, Models) and docs/MACHINES-FOR-DON.md. Don's framing: this jam is the origin story of the autonomous studio. Every method that works becomes a reusable studio job, not a one-off.
Rule one: nothing idles
The queue sends every job to tower 1 unless its spec says otherwise (run_on defaults to ["tower1"] in tools/queue/policy.py). So the laptop's 4090 and tower 2's 48 threads sat idle overnight. From now on every job spec names its machine on purpose: run_on: ["laptop"], ["tower1"], or machine: "cpu" with tower2_ok: true.
The three machines
| Machine | Jam role | Runs | Free models and tools it holds |
|---|---|---|---|
| Laptop, RTX 4090 16 GB, Windows | The stage: the only Unreal editor that builds, and the only compiler, so all packaging (code project, MetaHuman, LiveLink and ChaosClothAsset enabled) | UE 5.9 editor over MCP, packaging, Movie Render Queue for the film, ComfyUI when the editor is closed | ComfyUI (FLUX.2 klein, FLUX.1 schnell, TRELLIS.2, Pixal3D, Wan 2.2), Blender Cycles, Kimodo (no text encoder cached here yet), voices (Chatterbox, Kokoro, Dia, Qwen3-TTS, Maya1, VoxCPM2), Ollama (gpt-oss 20b, qwen3 14b, qwen3-vl 8b), OBS, Android SDK and JDKs, butler |
| Tower 1, RTX A5000 24 GB, Ubuntu | The factory and the critics: all generation overnight, and free judging | Queue jobs: images, 3D props, textures, music, voices, motion, video previz | ComfyUI (same models plus Wan 2.2 at 720p), ACE-Step 1.5, Kimodo SOMA (cleared for commercial use; the SMPL-X variant never), Chatterbox, Kokoro, Dia, Maya1, VoxCPM2, faster-whisper, Demucs, RIFE 4.9, Marigold, PBRify, Blender, Ollama (qwen3-vl 32b, qwen3.5 27b, gemma4 31b, qwen3-coder 30b), critics (speaker and accent ECAPA, audiobox-aesthetics, CLAP) |
| Tower 2, Threadripper 48 threads, W6600 (no CUDA), Windows | Second unit and QA: the test editor, CPU work, bot playtests | MetaHuman builds by script (signed in, MCP on port 8000), headless UE runs and frame captures, Blender CPU renders, bot playtests, CI, packaged-build smoke tests | UE 5.9 with MetaHuman content, Blender, Godot, OBS, VMware |
Free brains (no Claude tokens)
- qwen3-vl 32b on tower 1: a first-pass visual critic for renders, props and frames before any Claude reviewer sees them. Claude reviewers only see what passes.
- qwen3.5 27b on tower 1: bulk drafting (shot lists, alternative director lines, prop lists) that Claude then edits. Every player-facing word still passes
tells.pyand a Claude read. - faster-whisper and the takes check: word error rate, loudness and timing on every voice take.
- The motion gate (
lab/motiongen/gate): numbers decide which motion take beats the last.
Claude, by tier (ORG.md)
- Opus (this chat): direction, story, taste calls, review against STANDARDS.md, hard debugging (the hair bug).
- Sonnet workers: scoped build tickets with a check (cue gate, edit bay UI, import scripts).
- Haiku: web research, reading logs, checking outputs, the board's routine checks.
- Cheap API (Don's key, Haiku first): unattended batch judging once a worker script exists. A ticket for after 4 PM.
Pipeline for one take
- Story beat in
STORY.md, then a shooting script (prep/scripts/director/format) with cues and every fail branch. - Motion: Kimodo candidates on tower 1, the motion gate, retarget to the MetaHuman skeleton (twist the upper arm and thigh), joint-limit check.
- Set: props (TRELLIS.2 via ComfyUI), textures, Megascans kits we already hold for dressing, lights.
- Voice and sound: director and crew lines in Chatterbox or Dia, the takes check, then sound effects and score.
- Unreal on the laptop: Sequencer take, cue gate, fail branches, edit bay. Second-unit captures on tower 2.
- Critics: qwen3-vl first pass, then a fresh Claude reviewer against STANDARDS.md, then the board for Don.
Running now
- Tower 1: overnight batch 1 (music and voice done; props running; textures queued), then stunt motion batch 1 (23 moves at 3 seeds, seq 562, normal priority, so it runs before textures).
- Laptop: idle GPU. The Kimodo text encoder isn't cached here; copying it from tower 1 makes the laptop a second motion farm.
- Tower 2: test editor open, idle until the hair work at 4 PM.