DisCo: World Models with Discrete Camera Motion Control
Abstract Overview
DisCo is a controllable video world model that replaces continuous camera trajectories with a compact discrete action space for camera motion control. The paper argues that continuous camera representations create entangled conditioning features, which weakens reliable action following under complex motion sequences. The method discretizes camera motion into 27 translation-rotation primitives, injects these action embeddings into a video diffusion backbone, and distills the teacher into a four-step autoregressive student for efficient long-horizon generation. The authors also introduce DisCoBench, a benchmark covering short-term action following, dynamic scenes, and long-horizon generation, to evaluate visual quality, temporal consistency, and controllability across these settings.
Novelty
The main novelty is framing unreliable camera control in video world models as an action-representation separability problem and addressing it with discrete camera-motion primitives instead of continuous trajectory conditioning. The work is also distinctive in pairing this representation with an action-guided distillation pipeline for efficient long video generation and introducing DisCoBench for more systematic evaluation of controllability.
Results
On DisCo-ActionBench, DisCo attains the highest reported Action Score (0.6332), exceeding Matrix-Game2 (0.5880) while also leading the listed methods on several VBench++ quality measures such as subject consistency, background consistency, motion smoothness, and aesthetic quality. On DisCo-DynamicBench, it achieves the best FVD (220.51) and a competitive VQA score (0.5448), indicating strong performance on dynamic exploration videos. On DisCo-LongBench, it outperforms the compared baselines in consistency, motion smoothness, and aesthetic quality, and ablations show that discrete actions reduce feature similarity and improve action following over continuous representations.
Key Points
- DisCo converts continuous camera trajectories into 27 discrete translation-rotation action primitives to improve separability of control signals.
- The model combines an action-conditioned diffusion teacher with a four-step autoregressive student using self-forcing distillation, windowed attention, and an attention sink for long-horizon generation.
- Experiments and ablations on the newly introduced DisCoBench indicate stronger action reliability and competitive visual-temporal quality than the reported continuous- and discrete-action baselines.