Bud Knows Ball

The brief

Bud Knows Ball already had a mascot: a photoreal CGI silverback with an established brand voice — deadpan, unbothered, permanently talking about sport. The job was to put him into short-form vertical video in the style of the AI Bigfoot vlogs, without a studio, a shoot, or an animator on retainer.

The hard part

Character drift. Generate ten clips of the same character from text and you get ten slightly different characters — the face shifts, the fur changes, the brand stops being recognisable. That is the failure mode that kills AI mascot content, and it had to be solved before a single clip was worth publishing.

What we built

Two-stage generation pipeline — stage one produces a still keyframe from the existing brand art with an image model; stage two feeds that still in as the start frame for video. The face is locked before anything moves.
A character bible with a fixed identity block — one canonical description of Bud, pasted verbatim into every prompt so the model is never asked to reinvent him.
A format bible — shot grammar written down as rules rather than vibes: selfie POV at arm's length, wide-angle with mild barrel distortion, slightly low angle, constant handheld shake, available light only, 8 seconds, 9:16.
A scripted season — twelve clips written to the format, plus four topical ones, each with a ready-to-run prompt, so producing the next one is a copy-paste rather than a creative restart.
A production log — every generation recorded with its job ID and a verdict, so the prompts that work are known rather than remembered.

What happened

Five clips generated and delivered in the first session, with sixteen scripts written: a twelve-clip season and four topical clips. Bud looks like Bud in every one of them. Clips are produced from written scripts against a fixed character and a fixed shot grammar, so the next batch does not depend on rediscovering what worked last time.