CausalAlign: Causal Alignment of Video Generation Models for Consistent Ego-Motion

CausalAlign. Causal Alignment of Video Generation Models for Consistent Ego-Motion. By Kai Zhu and nine others, from Peking University, the Beijing Institute for General Artificial Intelligence, and three other institutions. Published on the twenty-first of October. Its code repository is worked out: about seventy source files, with a readme and a licence.

Diffusion-based video generation models can produce videos that look plausible, yet over long horizons the camera's ego-motion tends to drift. Ego-motion is how the viewpoint changes from one frame to the next, and it is what keeps geometry and trajectories consistent. When it drifts, the scene stops holding together as the camera moves, however good each frame looks on its own.

CausalAlign is a causal alignment framework with two components. The first is diagnostic: find out where and how precise ego-motion lives inside the model. The second is corrective: strengthen the ego-motion state at the places the diagnosis identified.

For the diagnosis, the authors use linear probes and causal interventions to study how ego-motion is represented, used, and retained within the video model. Probes ask whether the information is present in a layer's representation. Interventions ask whether the model actually depends on it. What they find is a rise-and-fall pattern during denoising: precise ego-motion is not held steadily through the process, and that pattern is what picks out the layers to target next.

For the correction, they align the representations at the identified layers to a 3D foundation model, and they do it through what they call delta-token space. That space encodes the changes between consecutive frames. Because it holds differences rather than frames, the alignment concentrates on the dynamics of ego-motion instead of on static scene content.

The base model is Wan2.2, fine-tuned on DL3DV and RealEstate10K. CausalAlign substantially improves motion accuracy, geometric consistency, and temporal smoothness, while preserving visual quality. As one example, geometric consistency improves by more than twenty percent relative to the base model. Better motion is obtained without trading away how the videos look.

The gains carry into embodied settings. Using the fine-tuned video model, CausalAlign raises the success rate of world action models by ten points on the mobile manipulation tasks of the RoboCasa365 benchmark.

The method follows a find-then-fix order. Probing and intervention locate where ego-motion is held during denoising, and alignment in a space of frame-to-frame changes reinforces it at exactly those layers. A steadier camera in generation then shows up downstream, as noticeable gains in embodied models.