video-engine

name	video-engine
description	Video generation pipeline guide for Synthex. Documents the 4-provider architecture (Runway, Synthesia, D-ID, Remotion), FFmpeg post-processing, media library integration, and God Mode gating.
metadata	{"author":"synthex","version":"1.0","engine":"synthex-ai-agency","type":"reference-skill","triggers":["video generation","video pipeline","remotion","ffmpeg","video rendering","runway","synthesia","d-id","avatar video","video processing","media generation"]}
context	fork

Video Engine — Video Generation Pipeline Guide

Purpose

Synthex generates videos through a multi-provider pipeline with 4 providers, FFmpeg post-processing, ElevenLabs voice integration, and a Supabase-backed media library. Remotion is God Mode only.

This skill documents the full pipeline, provider differences, and the architectural decisions at each stage.

Visual generation (binding): all images/video route through the grounded pipeline — see .claude/rules/real-images-only.md + the grounded-visuals skill. Direct provider calls fail CI.

⚠ SUBSTRATE REALITY (SYN-43 / SYN-48) — read before using any provider below. Runway ML, Synthesia and D-ID are NOT in the owned Synthex stack and are not sanctioned image/video substrates. The ONLY sanctioned entry point for image generation is generateImage()/generateBatch() in (the static guard test fails CI on any direct provider call outside the service layer). Video generation is grounded-by-default via — it seeds from owned references in (manifest.json, 143+ subjects incl. 135 CCW products) and the private Supabase bucket (customer job photos, ingest via ); no owned coverage ⇒ and the run is BLOCKED — "No owned references for this subject — add real photos to the reference library first." is the sole audited escape hatch (output stamped ). Trained LoRAs ( — , trigger ) auto-apply per industry; retrain via . Corpus growth from owned videos runs through the Railway media worker + MCP tools + frame-extraction scripts. MCP studio tools / inherit all these defaults.