Automated YouTube production

The same character. Every shot. Every episode.

Video-Ani turns a one-line premise into a finished, narrated, cinematic episode — script, shot plan, footage, dialogue, assembly, thumbnail and a Private YouTube upload — while an identity engine guarantees the cast never drifts. Approve a character once; the system reproduces exactly that face, outfit and voice in every future scene, or the shot never ships.

The operator dashboard is authentication-gated — it drives real generation spend. This page is the read-only public window into the platform.

Live platform stats

Aggregates read directly from the production system.

–
Characters created
–
Locked identities
–
Episodes submitted
–
Episodes completed
–
Tracked spend

Counters refresh every minute.

How an episode is made

One premise in, one uploaded episode out. Every stage persists its state, so a paused or failed run resumes exactly where it stopped.

Premise → script & shot plan

Claude writes the script, breaks it into scenes and shots, routes each shot to the right engine and prices the whole plan before a cent is spent.

Claude Opus

Footage generation

Veo renders principal cinematic shots, Runway supplies b-roll and transitions, Hedra performs lip-synced dialogue — every call carries the character's full reference package.

Veo · Runway · Hedra · ElevenLabs

Identity validation

Every shot is face-verified against the locked reference sheet and scored by a vision judge on clothing, hair, body and palette. Three strikes and the shot goes to human review — it never enters the edit.

InsightFace + vision judge

Assembly & publish

Shots are normalized and cut together, narration is mixed in, a validated thumbnail and SEO metadata are generated, and the episode uploads to YouTube as Private for final human sign-off.

ffmpeg + browser uploader

The identity engine

Character drift is treated as a build failure, not an aesthetic quibble.

Reference sheets, auto-built

From a short text brief the system generates design candidates, then a 20-image reference package — angles, expressions, full-body views, outfits, hair and facial-feature close-ups — each image validated against the approved seed before you ever see it.

Two approvals, then locked

You approve a design and a reference sheet; the identity locks. Locked versions are immutable — redesigns mint a new version while released episodes stay pinned to the version they were made with, forever.

Validation with teeth

Face-embedding similarity gates plus a structured vision rubric run on every generated shot. Failures regenerate up to three times, then park in a human review queue where you can approve, retry, or edit the shot spec.

Budget-capped by design

Each episode carries a hard spend cap. The plan is priced up front, every paid call checks the running ledger, and production pauses for explicit approval before a single dollar of overage.

Under the hood

A self-contained service on a k3s cluster — one pod runs the API, worker and state machine; a second, isolated pod holds the browser-driven YouTube uploader.

Python · FastAPISQLite (WAL) job queue Claude Opus 4.8Veo 3.1 (Gemini API) Runway Gen-4Hedra ElevenLabs TTSInsightFace ffmpegPatchright uploader k3s · Longhorn · Traefikcert-manager · CrowdSec