End-to-end character production pipeline for Zeramental — a dark fantasy AI series. Concept brief to finished shot sequence, fully automated. Character identity consistent across every frame.
Zeramental is a dark fantasy AI character series — WoW-style aesthetic, original characters, and a pipeline built to produce consistent cinematic output at scale.
The problem with AI character production at scale is identity drift. Run a character prompt 50 times and you get 50 variations. The pipeline here solves that: per-character LoRA models, dual-reference injection, and the SAME anchor mechanic enforce identity across every generation.
The output is film-grade. The process is automated. Brief goes in, CDN assets come out.
Character roster
Sin
ProtagonistFull-body blackout with pale veining and crackle. Face clear. Pale icy grey eyes. Chrome bracers and stiletto blade. Recurring male lead.
Jack
VillainMassive male figure. Black eyes. Chrome shoulder spikes. Primary antagonist. Shot in wide and mid compositions.
Kezia
Secondary villainFemale villain. Amber eyes. Blackout arms. Present in multi-character scenes with Jack and Sin.
Character brief
Physical attributes + visual identity written in photography language. No CGI vocabulary. Brief fed into the pipeline as structured parameters.
20-image reference build
Single-subject atmospheric shots across lighting conditions and angles. CSFD threshold check before training begins. 20 images is the floor — below this, LoRA underfit.
FAL LoRA training
Per-character model. $2.40–$8 per run. Three param schemes mapped. Training run checked against trigger word and config before production use.
Dual-reference injection
For multi-character scenes, both trained character models are injected as dual references. Standard single-ref fails when two trained characters share a generation.
Image generation — NB2 + GPT Image 2
NB2 (thinking_budget 1024) as primary. SAME anchor mechanic as mandatory closing line on every prompt. Photography-language vocabulary enforced throughout.
Video — Kling 3.0 + Seedance 2.0
Kling 3.0 for lip-sync and motion control. Seedance 2.0 for narrative multi-shot sequences. 180° rule, static camera priority, shot duration 6–9s target 8s.
QC + CDN delivery
CSFD threshold identity check. Duration and resolution verification. Auto-reject and re-queue on failure. R2 CDN upload on pass, URL registered in Supabase.
AI film and character production infrastructure is built. Tell me your project and what you need the system to produce.