From clips to worlds: where video generation is heading

The video-model field consolidated this autumn. In same-prompt testing published Oct 27, Wan 3.0 scored highest overall; Veo 3.1 Fast was the pick for expressive people in short clips with sound; Seedance 2.5 owns long takes (30 seconds, native audio, dozens of references); and Kling 4.0, launched Oct 15, leads on on-screen text and human motion. Sora is gone — OpenAI discontinued its apps in April 2026 — and the market barely blinked.

Look at the feature race and a direction shows: keyframes, first/last-frame control, video extension, subject references, draft modes. Every one of those is a control surface — a way for the director to pin the world down so frame 40 still agrees with frame 4. Video models are being dragged, feature by feature, toward consistency.

Consistency is exactly the world-model problem. A video model with perfect memory and frame-level input stops being a video model; it is a simulator with a camera. That is why the Genie line matters to anyone who only cares about "video": it is the same research programme viewed from the interactive end.

Practical take for my own pipeline: pick per shot, not per brand — establishing shots and interfaces from the text/motion leaders, people and dialogue from the expressive ones, and anything that must stay consistent across a scene from the reference-heavy models. And keep an eye on the seam: the first tool that makes a generated place hold still will change previs overnight.