WEBVTT

NOTE AI Engineering Visualised: https://understandingdata.com/ai-engineering-visualised/multi-shot-continuity/

1
00:00:00.267 --> 00:00:09.689
Every shot in a sequence is generated alone: the model never sees the shot before it, so nothing carries over unless your prompt carries it.

2
00:00:09.689 --> 00:00:22.142
Trace four shots written one at a time: gpt-image-1.5 cast a woman in a beanie, a man in a cap, then a stranger's hands, and the judge matched just 16 of 42 checks.

3
00:00:22.142 --> 00:00:34.375
Now paste one continuity sheet into every prompt, and the same woman holds all four shots, her cardigan measured mustard in each, but the sign becomes an easel and the latte black coffee.

4
00:00:34.375 --> 00:00:44.974
The non-obvious part: continuity lives in what you send, not in the model, so add the approved first frame as a reference, and the judge matched 59 of 60 checks.

5
00:00:44.974 --> 00:00:54.955
Watch the sign: the sheet never says where, so later shots put it on the counter, and the judge, comparing looks, passed it five times out of six.

6
00:00:54.955 --> 00:01:02.630
The fix: one state line per shot, sign on the door, latte in the mug, and that run matched all 54 checks.

7
00:01:02.630 --> 00:01:15.123
So treat continuity as data you carry, about 600 input tokens a shot: keep one sheet, send the approved frame, update the state after every shot, and check each cut, because the model remembers nothing.