How a clip is made
A clip is built in three quick passes, end-to-end in 60–90 seconds. The agent in the booth shows its work below — we like operators to know what they're paying for.
- 01
The still
A high-fidelity image model places your face into the chosen preset — a fan in the stands at a fictional baseball game. Identity preservation is the entire technical bet of the product; the agent's prompt has hard exclusions against any real team, league, or athlete. A safety classifier scores identity similarity and screens for trademark or logo bleed-through. If anything trips, the still is re-rendered up to three times. If it still trips, you aren't charged.
- 02
The motion
A diffusion video model animates the still — a slow camera push, a beat of you registering the camera, a small drift away. Five seconds, with native audio: stadium ambience, a brief crowd cheer, and the play-by-play line baked in.
- 03
The line
A separate strong language model writes the one line the commentator says, in the cadence of the era you chose. Under 14 words. Never your name. Never anything that quotes your optional context literally.
- 04
The graphic
A fictional BPN lower-third gets composited in: a faux network, a made-up team abbreviation pair, a faux score (always 3-2 or 4-1), an inning. The watermark — generated · catch.0p.studio — sits in the lower-right of the frame and as MP4 metadata. It survives cropping; it is part of the joke.
- 05
The finalize
A short ffmpeg pass burns in the graphic, embeds provenance metadata, generates a poster image, and writes the file to encrypted storage. The signed download URL is yours.
- 06
The cleanup
Your selfie is deleted after 30 days. The clip is re-downloadable for 90 days from your account. After that, the file is deleted from storage; the metadata stays for audit. We never use your face to train a model. The play-by-play voice is fictional and studio-owned.