82 Production Units: AI Video Failure Retrospective
If you only watch AI video demo reels, you'd think "this tech is almost there." But run a real project — novel to trailer, 44 shots broken into 82 production units — and you'll see the other side: the primary activity in AI video production isn't "generating." It's "discovering failure → diagnosing the cause → deciding whether to fix or abandon."
Key Numbers Up Front
| Metric | Value |
|---|---|
| Total Production Units | 82 |
| First-Frame (Keyframe) Passes | 39 (47.6%) |
| GPT-Image2 Calls | 430 |
| GPT-Image2 Failures | 99 (23%) |
| I2V Successful Calls | 38 calls, 220.1 seconds |
| I2V Cost | ¥198.09 |
| Agent Token Consumption | ~1 billion tokens (Codex 573M + Claude ~421M) |
| Total Project Cost | ¥2,234.67 |
"Less than half passed" isn't AI being bad — it's that real production quality standards are far stricter than demo showcases.
Five Failure Categories
82 PU failures fall into five categories, each demanding a different strategy:
| Type | Symptoms | Core Strategy |
|---|---|---|
| Character Consistency Failure | Same character looks different across shots | Anchor system + candidate caps + local fixes |
| Spatial / Scale Drift | Scene spatial relationships and scale drift | Spatial constraint keywords + scene anchor images |
| Multi-Character Collapse | 3+ people — spatial relationships uncontrollable | Split-and-composite route, or downgrade to suggestive shots |
| I2V-Stage Failure | First frame OK, but video jumps / flickers / deforms | Preflight Gate + candidate caps |
| Tool / Platform Failure | API failures, Agent drift | Cost log + task cards + manual gate |
Key Takeaway
The biggest cost driver isn't API pricing — it's rework. Every failure that makes it past the gate compounds downstream. The Preflight Gate (a 12-point checklist before any first-frame enters I2V) was the single highest-ROI investment in the entire project.
Full Chinese original with complete failure case breakdowns available at 82个生产单元的AI视频失败复盘