Skip to content
Case study · delivered to a client · built on GA

A production line
that learns from every episode.

From a paragraph of story text to a publishable vertical comic-drama episode: multi-agent storyboarding → GPU video generation with a premium tier → a judge · attribute · mutate closed loop → assembly. Each episode's retrospective becomes rules for the next.

17finished episodes across two series
16 minof delivered runtime
7.3average seven-dimension score, 10-point scale
193deterministic regression tests, all passing
The pipeline

Nine stages, one closed loop.

01

Script → story beats

A writer agent splits each episode into beats — hooks, twists, cliffhangers — and carries over the previous episode's end state.

LLM
02

Director's brief (GA harness)

A tool-using director agent checks the asset table, series rules and last episode's weak dimensions, drafts, self-checks against deterministic validators, then submits.

GA · tools
03

Shot cards (camera + sound)

Blocking, focal length, camera position, lighting, performance beats, line placement and continuity contracts; 40+ card-level validators.

LLM · validators
04

Visual director

A VLM reads the director's board — previous end frame, character anchors, scene plates — and writes continuity state and exit frames.

VLM
05

Engine routing

Identity-anchored, relay (previous end frame) or re-anchor generation; a premium tier only for key or rescue shots, with per-episode caps.

GPU · premium tier
06

Shot-level closed-loop search

Candidates in parallel → seven-dimension VLM judging plus objective signals → identity/costume audit → re-judge before acceptance → mutate only on hard faults, otherwise resample.

judge · search
07

Verification

Whole-chain re-scoring with per-series calibration; identity and costume gates cap drifting shots; cut points checked by objective signals.

verify
08

Assembly

Native voice kept line by line, voice cloning only for missing lines, per-face lip sync, adaptive transitions, ambience bed, BGM and subtitles.

assembly
09

Release & retrospective

Episode + report → lessons split into craft and ops → written back as series rules → the next episode's director starts with them.

self-evolve
Shot level: judge → attribute → mutate / resample → re-judgeEpisode level: weak dimensions + costume/cut audits → lessons → series rules → next directorCalibration level: zero-cost re-scoring tunes acceptance and weights; low temperature + re-judging suppress noise
Quality mechanisms

Judged like a critic, verified like an engineer.

Low-temperature judging + re-judging

Three judgments of the same clip went from σ 0.70 to 0.25; candidates are re-judged before acceptance and a third pass breaks large disagreements.

Identity and costume gates

Registered leads are compared frame by frame against approved designs; drift is a hard fault during search and a cap at verification.

Long-take verification

Shots of nine seconds or more are judged as a camera journey on a temporal strip rather than as isolated frames.

Adaptive transitions

Objective cut-point signals — histogram, luminance, motion on both sides — decide between hard cut, short dissolve or dip to black.

Native voice first

Lines are matched by phonetic similarity to keep the generated performance; cloning and per-face lip sync only fill the gaps.

Rules that route

Each retrospective is split into craft lessons that reach the director and ops lessons that stay out of creative prompts.

This production line was built for a client. Titles, footage, characters and per-episode data are under NDA and not shown here; the methodology and aggregate figures above are shared with permission. For a similar pipeline in your domain, get in touch.

Want a self-evolving pipeline for your content?

The harness, judges and evolution loop transfer to other media and other domains.