This week Tencent ARC open-sourced a method called SCoPE on Hugging Face: built on top of Alibaba's Wan2.2 video model, it layers in a "camera trajectory" as positional encoding. Given a first frame, a text description, and a camera path, the model outputs a finished clip with controllable camera work. In other words, AI video generation is moving from "gacha pulls" toward "shooting to a script."

What this is

SCoPE stands for Sightline-Coordinate Positional Encoding. What it does is concrete: inside an existing video diffusion model (an AI architecture that iteratively denoises random noise into video frames), it injects the camera position and movement direction as additional positional parameters.

One of the biggest pain points of AI video generation has always been the camera. The model wobbles on its own, and its moves follow no logic. If you want a steady shot, a deliberate push, pan, or tilt, you basically have to keep rolling the gacha — regenerating over and over until you land something usable. The core idea behind approaches like SCoPE is to take the "cinematographer" out of the output and put them into the input: you write a camera script, and the AI shoots to it.

The repo ships the full weights (model parameter files) of Wan2.2-I2V-A14B — a 14-billion-parameter image-to-video model — so you can download and run inference directly, no extra assembly required.

How the industry sees it

Controllable camera work is one of the main battlegrounds in video generation this year. Runway, Pika, Kuaishou Kling, and ByteDance's Jimeng are all pushing in similar directions; SCoPE is just one technical implementation path within the open-source community. The fact that Tencent ARC chose to stack on top of Alibaba's Wan2.2 rather than train its own base model tells us something: in the open-source ecosystem, whose foundation model is strongest matters more than whose brand is loudest.

Worth holding onto a few reservations. Camera-controllable does not mean storytelling-controllable. AI still doesn't know when to use a slow push to build tension, or when to cut hard for pacing — the grammar of the lens is something the model can't pick up. Another real-world issue is commercial licensing: the code typically ships under a commercially friendly license, but the model weights may carry additional terms. Enterprises need to read the fine print before deploying.

Impact on regular people

  • For enterprise IT and marketing: budgets for outsourcing short-video ads and product-introduction clips have more room to keep shrinking — a 5-second product shot no longer strictly requires a professional cinematographer.
  • For individual careers: self-media creators and short-video operators get another handy tool, but taste, scripting, and storyboarding remain scarce skills. The tool won't make content good on its own.
  • For the consumer market: regular users on Douyin and WeChat Channels will increasingly run into AI-generated content whose camera work "doesn't feel like AI" — nothing headline-grabbing, but it's already happening.