Week 10: Sustained coherence

Why keeping one visual language across linked pieces is harder than making any single piece look right, and what breaks first

26 April 2027

Open the slides

Case study: Nightshade & Nectar#

Nightshade & Nectar, a supernatural-romance slop serial posted in twice-weekly two-minute instalments, is a useful failure because it fails slowly. Episode one establishes its lead — amber-lit, a locked reference face, a recurring line of dialogue delivered in the same too-smooth cadence every episode — and by episode nine the lead’s face has drifted enough that long-time viewers comment on it before the plot. Nobody changed the seed on purpose; seed-locking degrades under small prompt edits, upstream model updates, and the ordinary entropy of a pipeline run by different operators on different days. Call this seed drift, and treat it as the sustained-coherence version of the six-fingered hand: not a single bad frame, but a slow divergence a single-clip critique would never catch, because a single-clip critique only ever sees one clip.

Where coherence actually lives#

Coherence lives in the parts of a production that don’t show up in any one frame: a locked colour grade applied as a fixed post-process rather than described fresh in every prompt, a reference-image set fed to every generation rather than a text description re-interpreted each time, a story bible that states the two or three rules everyone working on the piece has to obey regardless of which tool or operator touches a given shot. Nightshade & Nectar’s actual visual-language document, as far as it’s inferable from the output, seems to be “amber, slow zoom, and the line ‘I’ve been waiting’” — three rules, thin enough that the ninth episode can drift a face and still count as on-brand, because the bible never specified the face.

This is also where slop borrows most obviously from conventional serial television, which solved the same problem with style guides, standing sets and casting-continuity clauses decades before anyone automated it — the difference is that a slop pipeline’s failure mode is silent (the seed drifts and nobody notices until a viewer says something) where a human production’s failure mode is loud (an actor is visibly unavailable, a set burns down). Sustained coherence, in either case, is enforced by a document someone actually checks against, not by everyone involved sharing good taste.

This week’s set has you write that document for your own capstone, before the shooting — such as it is — starts.