An AI video studio built by one creator — with Claude as assistant and researcher, turning careful questions into proposals, research, and the hours of engineering behind every release.
The method, on repeat — one strong idea, one pipeline, one iteration. Improved and expanded through research alone.
No team, no roadmap theatre — a single, repeatable rhythm. The creator asks; research answers; one pipeline gets one honest iteration better.
The creator poses precise, testable questions — the kind that move one thing forward.
Claude researches, weighs options, and drafts proposals — the hours of legwork, on tap.
A single idea becomes a single pipeline. No sprawl, no half-finished forks.
Each change is benched and looked at. It ships, or it goes in the failed column — honestly.
Every Lumen clip is rendered on a T4 — a graphics chip that lives in a data center, and one Lumen borrows for free. Here's the gist, in plain terms.
An NVIDIA T4 — a GPU built for inference (running trained models, not training them). 16 GB of memory, 320 Tensor Cores, and it sips just 70 watts.
Data centers stack rows of them in servers, wired together with high-speed links (PCIe / NVLink) — quiet aisles of compute you never see.
Lumen owns none of it. It borrows a single T4, for free, through Google Colab — and that one chip is what turns your prompt into video.
Simplified for the gist. Technically the T4 is an NVIDIA chip that Google (and others) offer in the cloud — not a "Google chip." And that 16 GB of memory is Lumen's ceiling: it's the exact reason clips render at 480p and not 4K — the whole model has to fit inside it.
These stay blank on purpose. A reel goes up only when a clip earns it — looked at frame by frame, not just measured. Placeholder until then.
Real questions pulled from the build log — some answered, some still open. This is the frontier the studio is working, in plain terms.
Can hidden physics priors — Verlet dynamics, momentum, restitution — steer a photoreal model's motion while never showing geometry?→ shipped as a hidden conditioning layer; realism intact.
Why does the 16 GB memory wall behave identically on a T4 GPU and a v5e‑1 TPU — and what does that say about the real bottleneck in video diffusion?→ it's memory, not compute. Same wall, different silicon.
A sequential Mixture‑of‑Experts lets a "27 B" model run on a card that shouldn't fit it. How low can the active footprint really go?→ open — one 14 B expert resident at a time fits 16 GB.
Is CFG distillation genuinely lossless — two guidance passes collapsed into one at no quality cost?→ ~lossless on the low‑noise pass; the high‑noise pass still needs guidance.
Where is the "fuzziness floor" — the fewest inference steps before video breaks down?→ ~4 steps with distillation; below that it turns to mush.
Late‑clip face smear — is it lighting noise, or the image‑to‑video chain compounding softness across hops?→ the chain. Denoise made it worse, not better.
Where does a clip's "old‑film" character come from — the model, or a highlight rolloff that accidentally mimics film's shoulder?→ both; the rolloff is a controllable style, not a defect.
Most studios show only the wins. The failed column is the point: it's what a research‑first process actually looks like.