A nine-phase production recipe that takes a character brief through LoRA training, image generation, video sequencing, and automated delivery. No manual steps after intake. No rerenders from bad blocking.
Character brief
Physical description, role, tone, recurring visual identity. Written in photography language — no CGI or production vocabulary.
Reference set build
20-image bundle per character. Mix of single-subject atmospheric shots. Tested against CSFD thresholds before training.
LoRA training
FAL trainer — $2.40–$8 per run. Per-character models. Dual-reference injection for scenes with multiple trained characters.
Shot list + blocking spec
180° rule enforced. OTS, static camera for speed. Axis breaks, eyeline matching. Shot duration: single-shot 6–9s, target 8s.
Image generation (hero frames)
NB2 (thinking_budget 1024) + GPT Image 2 as primary models. SAME anchor mechanic as closing line for consistent lighting per character.
Video generation
Kling 3.0 for lip-sync + motion control. Seedance 2.0 for narrative sequences. Multi-shot syntax with per-segment timing brackets.
QC pass
Duration, resolution, character identity check. CSFD threshold: frames must score above threshold for character consistency to pass.
CDN delivery
Cloudflare R2 upload. URL registered in Supabase. Available to downstream consumers immediately.
Pipeline loop
n8n orchestrates the full chain. Brief in, CDN URL out. No manual steps after intake.
180° rule
Characters maintain consistent screen direction across cuts. Breaking it reads as a jump cut. Hard rule — not a guideline.
Static camera priority
Static setup generates faster and holds consistency better than camera movement. Only add motion when it serves the shot.
OTS > POV
Over-the-shoulder shots preserve character identity and make eyeline matching easier. POV loses the anchor character.
Shot duration 6–9s
Single-shot target: 8s. Short enough to hold consistency. Long enough to cut from. Multi-shot only when Kyle explicitly requests it.
SAME anchor
Mandatory closing line in every image prompt. Without it, per-character lighting is overridden by global scene logic. Breaks consistency.
Photography language only
No CGI, gaming, or production/BTS vocabulary in prompts. Strip production terms — models vocalize them instead of rendering them.
Other machines