AI video, but with a coordinator. And a team of specialists.
A small Director ANN plans the shot list. Then Camera, Motion, Depth, FX, Audio, and Quality ANNs do their part. Each is published, versioned, and rateable in the Aigarth marketplace. This page is the build-in-public view of the proposal.
The team
Seven small, specialized ANNs collaborate to produce a video. Each one excellent at exactly one part of the job.
Director
Reads the prompt, plans the shot list. Returns structured JSON describing scenes, durations, camera, audio, and style.
Camera
Picks the cinematic move. Slow push in? Dolly left? Whip pan? Knows 50 ways to move a camera and when to use each.
Motion
Generates in-between frames from a keyframe and a camera move. Pure interpolation, deterministic, fast on CPU.
Depth
Foreground / background separation for parallax and depth-of-field. Uses MiDaS on CPU.
FX
Particles, steam, light glows, atmosphere. Domain-specific. A Steam FX ANN and a Rain FX ANN can coexist.
Audio
Aligns narration, music, and sound effects to the timeline. Knows when to dip the music under a voice.
Quality
Watches the result. Spots artifacts, weird motion, dropped frames. Decides whether to ship or to re-run a stage.
How a render flows
A user prompt becomes a shot list. Each stage consumes it and produces an artifact. The Quality ANN at the end decides whether to ship or to re-run a stage.
Every stage emits a structured event. A bad stage can be re-run alone. The rest of the pipeline is preserved.
Will it fit the Aigarth architecture?
We evaluated four dimensions. Most of the platform is ready. The continuous-learning pillar is greenfield, and the most important one.
ANN registry, versioning, reviews, marketplace listings, deploy hook: all ready. Add `role` enum, `ann_pipelines`, `ann_pipeline_runs`.
Typed jobs, clusters, regions, reservations: all ready. Add `compute_workers` table, worker protocol, and `/v1/workers/*` namespace.
No per-ANN quality signals today. Add `ann_feedback_events` table and a `ann_version_metrics` materialized view + cron.
On-chain ANN ownership, worker registry, and reputation map cleanly. Blocked on real K12 signature verification.
Roadmap
Five phases. Re-evaluate at the end of Phase 1 with a real 30-second render before committing to Phase 2.
Architecture evaluation, blog article, dashboard, tracker. No code in production.
Single Python process. A prompt in, an MP4 out, every stage recorded. Re-evaluation gate at the end.
Publish ANNs as real listings. Sell pre-composed pipelines. Wire up per-version metrics and reviews.
Worker registry, worker protocol, `aigarth/worker-video` image. Real per-stage cost.
On-chain ANN ownership, worker registry, reputation, rewards. Off-chain rollups committed to chain.
What it costs (illustrative)
30-second 1080p render on a single CPU worker. Numbers are placeholders that will change as we ship and measure.
| Stage | Cost (QU) | Latency |
|---|---|---|
| Director (LLM call) | 0.05 | 2 s |
| Camera (rule-based) | 0.01 | < 1 s |
| Motion (OpenCV, CPU) | 0.50 | 60 s |
| Depth (MiDaS, CPU) | 0.30 | 45 s |
| FX (compositing) | 0.10 | 10 s |
| Audio (alignment) | 0.05 | 5 s |
| Quality (model) | 0.10 | 15 s |
| Render (FFmpeg) | 0.05 | 30 s |
| Total | ~1.16 | ~3 min |
Per-stage pre-flight estimates let the user decide: pay for a faster worker, swap in a cheaper ANN, or skip the stage entirely.
The single hard precondition
Everything in Phase 4 (the on-chain portion) depends on one thing we haven't shipped yet.
services/qubic is currently format-only. The address regex passes, but the cryptographic signature isn't actually checked. This blocks every on-chain write in the proposal. The fix is to use @noble/curves (already a dep) for real K12 verification. ~3 story points of focused work, no new architecture.
Follow along
This is a build-in-public project. The dashboard at localhost:4000/video is the source of truth for what is real today.
Numbers, phases, and product surfaces in this article are illustrative placeholders that will change as we ship and measure. Aigarth Cloud is a build-in-public project; the dashboard is the source of truth.