What 1,003 paid AI video runs taught me about failure modes

I paid for 1,003 generation credits across six AI video templates before charging anyone. Not demo credits — real runs, including the ones that failed.

The first thing I learned is that the model is rarely the bottleneck. Scope is.

1. Copyrighted audio is a wall, not a risk

A rap-duo template took three attempts. Every run that referenced the copyrighted song was rejected by the provider's content filter, including a silent render tied to the same performance. Dropping the copyrighted audio entirely made it pass. The fix was product scope: ship the template silent, let users attach the trending sound in TikTok themselves. 510 credits to learn that no amount of prompt engineering beats a model-provider policy.

2. Motion references are dependencies

Scenes with one central performer survive a face swap. Multi-person scenes fight the model. That is not a tuning problem; it is a template-selection problem. If the reference has two people, expect a higher failure rate and price the retry into the run.

3. Photo quality decides most of the outcome

Front-facing, well-lit, both hands visible. The difference between a usable clip and a warped one is usually the input photo, not the model. We moved that guidance onto every template page instead of writing a better prompt.

The product that came out of this is deliberately narrow: seven templates, one photo in, a 7–21 second vertical clip out, one-time packs from $9.99, no subscription, failed renders refunded.

The newest template is the same lesson in a smaller package — Bicep Heart AI: one photo becomes a ~7 second 9:16 clip of you smiling and flexing a heart on your bicep. Same flow, same refund rule, no sign-up.

If you are building in this space, the useful question is not “which model?” but “which failure mode am I pricing in?”