Thinking · Newsletter
What Veo3 is actually good at, and what it still cannot do
A year of running generative video in production, written down honestly rather than as a demo reel.
This piece is now part of a longer guide: UGC ads for performance: the complete guide.
Most writing about generative video is either a launch post or a eulogy. This is neither. It is what we found after putting these tools into real production schedules.
Genuinely good
Single-speaker monologue. Lipsync and emotional range on one person talking to camera is the strongest thing the current models do, by some distance. Prompt it well and the delivery holds up next to a shot creator.
Camera movement. Pans and tilts come out cinematic with almost no coaxing. Getting the same move from a phone in someone's kitchen costs a tripod, a retake and patience.
The fifteenth variation. Once a concept works, generating another framing of it is close to free. That used to be the part of production that ate the schedule.
Still not there
Hands and products. Anything held, opened, applied or worn. It is improving and it is not reliable.
Multi-person scenes. Two people interacting falls apart in ways that are hard to predict and expensive to discover late.
Continuity across shots. Keeping a character consistent over a sequence still takes more prompt engineering than it takes to just film it.
The practical rule we use
Generative for the parts that are expensive to vary. Human for the parts that have to be believed. Anything in between, test both and let the account decide.