Character development end-to-end: concept brief to LoRA-trained model to consistent multi-shot generation. The system keeps identity stable across lighting changes, shot types, and generation runs.
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.
Reference set build
20-image bundle per character. Single-subject atmospheric shots across lighting conditions, angles, and distances. CSFD threshold check before training.
20 images is the data-backed floor — below this, LoRA underfit.
LoRA training
FAL LoRA trainers. $2.40–$8 per run depending on quality target. Per-character model. Training parameters verified against three FAL param schemes.
Three param schemes mapped — each has different trigger words and training config.
Dual-reference injection
For scenes with multiple trained characters, dual-reference injection ensures both identities are preserved in the same frame without blending.
Standard single-ref fails when two trained characters share a generation.
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.
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.
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.
Other machines