The expensive part of AI video isn't generating it, it's regenerating it
Khalizo · reddit · 2026-10-09
A developer building a product on top of video generation models shares months of hard-won lessons: the per-clip price barely matters — the real cost is the regenerate loop. A 30-second video with 6-8 shots can easily burn 40 generations because each shot can fail independently.
Key takeaways:
- Cheap previews beat clever prompts: validate character, color and composition with stills first, at a fraction of the cost of motion.
- The planning model matters more than the video model: cutting the planning step's "thinking" thinned plans and caused failures — the shot-planning step decides quality.
- Consistency is a product problem: models have no cross-shot memory; reference images plus a written brief per shot beat prompt tricks.
He also learned that an upstream capacity incident (a step taking 60+ seconds against a 29-second gateway timeout) turns "the model is slow today" into a product outage unless designed for. His product Melodious turns songs into music videos with a free first storyboard; he asks how others handle the retry-cost problem.
More from coding & agent
- Dev is vibe coding an open-world RPG for the N64 — Dimillian · 2026-10-09
- Rule of thumb: automate predictable workflows, reserve AI agents for decisions — sujingshen · 2026-10-09
- Shakespeare: an open-source Mac typing-stats app built with Claude Code — whitew0lf · 2026-10-09
- OhMyGame open-sourced: an AI agent builds games and then plays them — sujingshen · 2026-10-09
- Codex turns vague install instructions into real charges once a card is stored — CurieuxExplorer · 2026-10-09
- LLM Wikis: Let an AI Curator Build Growing Knowledge Bases in Your Obsidian Vault — dSebastien · 2026-10-09