Infinity Engine

Project links and history

• First substantive build: 20 July 2026.
• GitHub repository: infinity-engine.
• Public site: visit the public site.

Related public projects

Each link below reflects an evidenced family, lineage or direct connection. This project has 4 relevant public connections.

i C. infinity music and content system

• explorer-music-video-lab - public page - later build; explorer-music-video-lab is earlier, explicit cross-reference, ordered build lineage, shared named build family.
• i-C-infinity - later build; i-C-infinity is earlier, ordered build lineage, shared named build family.
• i-C-infinity-music-universe - public page - later build; i-C-infinity-music-universe is earlier, explicit cross-reference, ordered build lineage, shared named build family.
• infinity-content-engine - public page - later build; infinity-content-engine is earlier, explicit cross-reference, ordered build lineage, shared named build family.

A model-agnostic pipeline that turns i C. infinity songs into visuals: lyric videos, comics, vertical micro-dramas, avatar course videos, and album aggregates. Cheap text thinking first, human direction points, expensive generation last.

Luke × Claude.

How it works

1. Vault (vault/, gitignored): one markdown note per song with YAML frontmatter against a fixed ontology (ontology.yaml). Built by ingest from the album-pack repo and/or a private local lyrics folder. Plain files and git, no app lock-in.
2. Ideation: brief writes a self-contained prompt packet per song. Run it through any LLM (or --run-claude to use the claude CLI directly) and merge folds the JSON answer back into the note. Status moves ingested → analysed.
3. Human gate: Luke edits the note, picks a direction, sets status to briefed.
4. Generation: jobs (comic panels, keyframes, video shots, TTS, lip-sync) are folders of spec.json + assets. A runner executes them: local box, rented GPU pod over SSH, or a hosted API. Which model and which host is pure config in providers.yaml. Runs are logged to SQLite.
5. Review: dashboard renders a static HTML view of the whole catalogue and where every song sits.

The knowledge base and orchestration stay on the trusted machine. Only per-job payloads travel to GPU workers, which are treated as stateless and untrusted.

Setup

pip install -r requirements.txt
python -m engine ingest
python -m engine status
python -m engine dashboard

Point config.yaml at your sources. On the production machine, set sources.local_lyrics to the cleaned lyrics folder.

Commands

Command	Does
python -m engine ingest	build/refresh the vault from sources (safe to re-run)
python -m engine status	counts by album and pipeline status
python -m engine validate	check every note against the ontology
python -m engine brief SLUG / --all	write ideation prompt packets
python -m engine brief SLUG --run-claude	run packet through claude CLI and merge
python -m engine merge SLUG FILE	merge an analysis JSON by hand
python -m engine jobs	recent generation job log
python -m engine dashboard	regenerate dashboard.html

Model and compute choices

providers.yaml holds the current registry (researched 2026-07-20): draft/standard/premium tiers for video, lip-sync, TTS, image, LLM and embeddings, plus where to rent GPUs and what it costs. Swapping a model or vendor is a config edit, never a code change. Licence notes are inline because the universe is commercial: everything in the default path is Apache 2.0 or MIT.

How to actually drive it, stage by stage, lives in the operator guide. The deeper plan lives in docs/PLAN.md.