Tencent's SCoPE nails AI camera control with tiny overhead

If you have ever tried directing an AI video model, you know the frustration. You ask for a smooth dolly forward, and the model behaves like a drunk drone pilot. Even worse, if the camera pans away and returns, the original scene usually morphs into something completely unrecognizable.

This week, Tencent researchers introduced SCoPE, short for Sightline-Coordinate Positional Encoding. Built for video diffusion transformers, the method achieves strict camera trajectory control while adding less than 0.1% parameters to the base architecture.

Giving every token a second address

Controlling virtual cameras in diffusion models has always been tricky because the network does not know where physical objects exist before generation begins. Forcing camera coordinates into the pipeline often breaks visual fidelity.

SCoPE tackles this by assigning each visual token its camera sightline ray as a second positional address. The user defines the trajectory, and the model locks that ray geometry directly into its attention queries and keys.

The team implemented this through a minimal Normalize-Gate-Inject mechanism. It supports both metric and relative camera poses while keeping the model's original RoPE encoding bit-exact. Despite adding under 0.1% extra weights, the geometry upgrade improved both controllability and visual fidelity across 5B and 14B model scales.

The 90-degree turn test

In practice, feeding SCoPE a single frame and a target trajectory reliably generates matching motion. The researchers showcased pullback-and-rise, push sweeps, S-curve reveals, crane-up shots, and dolly zooms. They even hooked up WASD keyboard inputs to steer third-person views in real time.

What stands out most is revisit consistency. In one test, the camera turns 90 degrees, moves through a gateway, and swings back. The original room remains rock solid instead of drifting into an hallucinated layout.

Crucially, this stability is not powered by an expensive external memory bank. It emerges naturally because the returning camera view reuses the exact same ray coordinates. The code is already out on GitHub under an Apache 2 license with local installation instructions.

Why it matters

Camera choreography is the backbone of filmmaking. Until now, creators had to prompt and pray that an AI generator would stumble upon the right cinematic motion. Having deterministic, frame-accurate camera control bridges the gap between viral party tricks and real production storyboarding.

Technically, Tencent's surgical approach is refreshing. While many labs slap bulky control networks onto transformers or retrain giant models from scratch, SCoPE acts as a plug-in geometric coordinate fix that costs almost zero extra compute.

That said, high-end production pipelines will want to see how it handles fast-moving human subjects and multi-shot transitions over longer clips. But for developers and studios looking to tame wild AI cameras right now, this is one of the cleanest architectural solutions we have seen.