Generative video is not a hypothetical anymore. It is already shipping in commercials, music videos, social cuts, and the occasional documentary B-roll. The honest question for a post-production studio in 2026 is not whether to use it. It is which model, in which shot, for which client, with which fail mode you are willing to absorb.
There is a comparison-table version of this conversation. You can find it on every blog. We are skipping it. What follows is shorter and more useful: where each tool actually earns a place in a working edit, and the three failures that all of them still share.
Where each model earns its keep
Sora 2 wins on photorealism. The physics is more coherent than the rest of the field — water behaves like water, fabric behaves like fabric, faces hold under closeups in a way that older models could not approach. It is the only major model to ship with synchronized dialogue and sound effect generation in the same pass. For establishing shots that do not need to integrate with live-action coverage, it is the strongest single image-to-video pipeline on the market.
Runway Gen-4 wins on consistency. Temporal coherence and camera control are the most conservative in the field — boring praise, but it is what professional editorial actually needs. When a director asks for a specific dolly-in or a slow truck across a static plate, Runway is the only platform that reliably delivers what was asked for instead of what the model thought looked nice.
Veo 3.1 wins on integration. Synchronized audio in a single pass — ambient, dialogue, sound effects all generated alongside the picture rather than added later. For social work where a ten-second cut needs to be three rounds of revisions deep by Tuesday, the round-trip economics matter as much as the picture quality.
Pika wins on style. It is not trying to fool anyone into thinking the output is photographic. For animated explainers, music videos with overt graphic identity, and stylized brand work, Pika delivers a distinct visual language faster than the others can match the brief.
Kling sits in the middle and wins on availability. It is the model most teams actually have access to without a waitlist. For producers who need to test an idea this afternoon rather than next week, that matters.
The three failures every model still has
Continuity is the first one. Generate two shots from the same prompt and they will not match — not lighting, not lens character, not the texture of the wall in the background. Cut between them and the audience feels the seam. The shorthand inside our edit bay is AI seams. We have learned to either commit fully to a stylized look that hides them, or use generative shots only as bookends rather than within a sequence.
Character consistency is the second. None of the major models ship a workflow that holds an actor's face across cuts the way live-action coverage does. Reference-image conditioning helps. It does not solve. The best current approach is to lock generative work to non-character coverage — environments, abstractions, atmospherics — and shoot real talent live.
Lighting drift is the third. Generate a daytime exterior. Now generate a continuation of the same scene. The sun has moved, the bounce on the talent's face has shifted, the color temperature is no longer the same. A grade can pull these closer. It cannot make them match.
“AI is a coverage tool, not an edit tool. The cut is still ours.”
The honest workflow
Generative B-roll for non-character work: yes, increasingly often. Generative title sequences and abstract transitions: yes. Generative establishing shots for environments we cannot shoot: yes. Generative sequences that need to integrate with principal photography: very rarely, and only with a heavy stylistic frame around them to hide the seams.
Every generated output gets the same metadata audit a piece of stock footage would get — provenance, license terms, model used, prompt archived. The IPTC standard added formal AI-provenance fields in its 2025.1 release. We use them. Auditors increasingly ask for them. So do platforms.
The marketing pitch from every generative tool is that it replaces the production budget. The reality, after eighteen months of testing them on real client work, is more boring. They are very good supplementary tools. They are not yet authors. The cut is still made by a human who decides which generated frame ships and which goes back.
