ad-creative-engine
Ad Creative Engine
One pipeline for paid-social creative: a winning static goes in, a Meta-ready testing matrix comes out — copy variations, image variations, and the best statics animated into video. Everything runs through the fal.ai API — one key, one pay-as-you-go bill. Statics default to GPT Image 2 (fal-ai/gpt-image-2/edit for reference-based variations — best-in-class text rendering for on-image copy); video uses Kling for cheap motion or Seedance for audio/UGC feel. Everything is quoted before it runs, capped by a budget, saved locally, and logged.
This replaces the subscription tools (Higgsfield, Krea, Freepik AI) with pay-per-asset generation — but the real edge is context: subscription tools can't read the brand's DNA file, banned-words list, or ICP voice-of-customer language. This skill won't write a line of copy without them.
The three modes
Matrix (the default when the user has a winning ad, hooks, or a brief): copy variations + matching static image variations. Requires the brand foundation files and a reference — this mode is a remix machine, never a from-scratch generator.
Animate: take chosen statics (usually matrix winners) and turn them into short video ads via fal.ai. No foundation files required if the user just says "animate these two."
Direct: one-off generation requests — product b-roll, background plates, a hero shot to comp into Figma. Route to the cheapest model that meets the bar (see references/models.md).
Modes chain naturally: matrix → user picks winners → animate. Never animate before the user has picked — don't pay to animate losers.