AI Production/Character Creation
Character Creation

Consistent identity.
At any scale.

Character development end-to-end: concept brief to character sheet with turnarounds to consistent multi-shot generation. The system keeps identity stable across lighting changes, shot types, and generation runs.

20
Ref images per character
Sheets
Multi-angle turnarounds
No training
Used as direct reference
Dual-ref
Multi-char injection
Character A — full-body reference
Character A — close-up, different lighting
Character B — OTS with Character A
Character A — video frame extract
Development pipeline

Concept brief to production-ready character.

01

Character brief

Physical attributes, tone, recurring visual identity. Written in photography language only — CGI and production terms stripped before any generation.

Photography vocab = no rendering artifacts from production language.

02

Character sheet + turnaround build

20-image bundle per character. Multi-angle turnaround set — front, three-quarter, profile, expression variants — used directly as reference. No training step, no fine-tuning run.

Sheet-based reference, not a trained model — higher consistency hit rate in production.

03

Dual-reference injection

For scenes with multiple characters, dual-reference injection ensures both identities are preserved in the same frame without blending.

Standard single-ref fails when two characters share a generation.

04

SAME anchor generation

Every image prompt closes with the SAME anchor mechanic. Without it, per-character lighting is overridden by global scene logic — identity breaks between frames.

Confirmed May 2026 — SAME anchor is mandatory. Not optional.

05

QC + delivery

Character identity check per output. CSFD threshold scoring. Auto-reject and re-queue if consistency drops. CDN delivery on pass.

QC gate prevents character drift accumulating across large batches.

In production

Zeramental — dark fantasy AI production project.

This pipeline runs in active production on Zeramental, a dark fantasy AI character series. Multiple trained characters, a locked visual style, and shot-consistent multi-model generation across every frame.

Per-character sheets with multi-angle turnarounds, versioned
Dual-reference injection for multi-character scenes
NB2 + GPT Image 2 as primary generation models
Kling 3.0 + Seedance 2.0 for video output
SAME anchor mechanic enforced on every image prompt
See the film pipeline case study →
Zeramental — Sin (protagonist)
Zeramental — Jack (villain)
Zeramental — Kezia
Zeramental — multi-character scene