Should AI shot selection prioritize visual impressiveness or narrative continuity?

Narrative continuity should win — every time. Choose the AI generation that connects your previous shot to your next one, not the one that looks most spectacular in isolation. A visually impressive shot that breaks story logic reads as fake to viewers; a continuity-matched shot makes the whole sequence believable, including the AI footage inside it.

Select AI shots by asking one question: does this generation make the previous shot and the next shot feel like one continuous world? In one documented production, the filmmaker chose a shoreline wide over more dramatic ocean generations specifically because the practical footage showed the actor on hands and knees emerging from water — the shoreline matched that body position and let the two scenes cut together seamlessly. The prettier options lost because they ruptured the match.

Continuity also tells you which shots to give AI in the first place. Wide environment shots are where AI is strong; close-up character acting is where it collapses — if the shot is close enough that the character has to perform, the footage looks fake and the story break is instant. Pull out rather than push in: use AI for the wides that place your action in a larger world (that same production built a full ocean environment around a 6x6 ft studio pool this way), and keep close-up performance on real actors.

Enforce continuity across generations, don't hope for it. Upload screenshots from your own footage as references so color, contrast, and look stay anchored to the practical material — reference-based generation produces dramatically more realistic matches than text-only prompting. Add environmental continuity cues explicitly, like matching haze between an exterior and the interior it cuts to; without them, AI scene transitions stop feeling believable. When a generation is close but not matching, pin it and request variations — "create another one like it, but change X, Y, and Z" — instead of starting over. Establishing the look and tone on the first shot of a scene also pays off downstream: in that production, once the visual language was set, the second shot in the scene generated correctly on the very first attempt. Tools like the invideo agent handle this as a conversation — it holds your references and prior generations in context and routes each shot to the right model, with Veo, Kling, and Seedance 2.0 all available in one place (Seedance 2.0 produced the wide ocean environments in the production above).

Visual impressiveness still has a job — in service of the story. A generation that looks striking but artificial at normal speed can work as an insert or cutaway: speed-ramp it 1,000–5,000% in your editor and splice the fragments into a short burst of energy. Then blend everything with a color grade and sound design — sound is the multiplier that makes AI inserts feel real. The point is role assignment: spectacular-but-fragile clips become inserts, continuity-matched clips carry the narrative.

Watch some of these to see what works for you:

story is king. The reason why this works is because if you look at the screenshot, it looks like I'm on my hands and knees and I'm coming out of the water. Making him on a shoreline would work the best because they go from one scene to the next and it feels like it matches well.

— Alex Arfaoui, filmmaker