Technology
Consistency is infrastructure, not luck.
The pipeline that makes a second season possible: versioned canon, locked identities, evaluated models and a delivery system that fans one master out to every platform.
The pipeline
Seven stages, one source of truth.
- 01
Development
Story structure is authored as versioned documents, not chat logs. Beats, canon and character bibles live in source control so every downstream stage reads from one truth.
- Canon registry
- Beat graph
- Narrative QA
- 02
Identity & Consistency
Each character gets a reference set, an identity lock and a continuity sheet. A drift monitor scores every generated shot against the lock and flags the ones that wander.
- Reference sets
- Identity locks
- Drift monitor
- 03
Model Training
Datasets are curated, tagged and versioned in-house. LoRAs and fine-tunes are trained on our own hardware against a benchmark set, scored, and promoted into a registry only when they beat the incumbent.
- Dataset curation
- LoRA training
- Eval harness
- Model registry
- 04
Generation & Direction
Shots are generated in passes against a shot list, with lens, blocking and coverage specified up front. Dailies are reviewed by a director, not sorted by a score.
- Shot list engine
- Batch generation
- Dailies review
- 05
Audio
Voice, score, dialogue and foley are produced against picture, then mixed and mastered to platform loudness targets with stems retained for every future version.
- Voice design
- Score
- Mix & master
- 06
Finishing
Compositing, cleanup, simulation and grade. Everything is conformed at 4K with an HDR pass, and the master is archived alongside the project that produced it.
- Compositing
- Simulation
- Colour grade
- Conform
- 07
Delivery & Automation
One master fans out into every required aspect ratio, language and platform spec, with publishing automation and an analytics loop feeding the next cycle.
- Versioning
- Localisation
- Publishing automation
Principles
How we decide what to build.
Self-hosted by design
Generation runs on open-weight models we train and host ourselves. The pipeline is not welded to one architecture — new models are adopted as the frontier moves — but they run on our infrastructure, never as a call to someone else's API.
Consistency is infrastructure
Character identity, set continuity and brand canon are enforced by systems, not by remembering. That is what makes a second season possible.
Human direction, machine throughput
Directors, editors and sound designers make the calls. The machines handle volume. We have never shipped a cut no one watched end to end.
Private models, not shared ones
Every client model is trained in isolation on their own material. Nothing is pooled across accounts, and no client's work is ever used to improve a model another client will use.
Provenance and consent
Training data is licensed or client-owned, voice cloning is consented and contracted, and every delivered asset carries a provenance record. Client material is never fed into a general-purpose model.
Stack
What sits behind each stage.
Every capability below runs on infrastructure we control. Model architectures change as the frontier moves; what does not change is that client material is processed in-house rather than sent to a third-party service.
Generation
- Diffusion video models
- Image models
- Custom LoRAs
Consistency
- Identity locks
- Reference sets
- Drift scoring
Training
- Dataset tooling
- LoRA / fine-tune
- Eval harness
Audio
- Voice synthesis
- Score tooling
- Mix & mastering
Finishing
- Node compositing
- Simulation
- Colour science
Delivery
- Transcode farm
- Localisation
- Publishing API
Start at 9am
Tell us what you want to make.
Send a brief, a script, a reference board or a sentence. We will come back with an approach, a schedule and a number.