CamRM: Towards Reliable Camera-Motion Control in Video Generation with Motion-Grounded Rating Models

Test-time scaling provides a practical way to improve camera-motion control in video generation: sample a clip, rate its alignment with the request, and revise the prompt if needed. This process requires a reliable rating model (RM) to determine when to stop. Geometry-based methods can struggle when object motion dominates the view, while fine-tuned visual language models (VLMs) often fail to adapt to synthetic videos. Our benchmark of $1,064$ human-labelled real and synthetic videos shows that these methods struggle to judge camera-motion alignment. Strong performance on real video does not ensure reliable judgment on generated clips: a model tuned only on real footage reaches $82.3$% balanced accuracy on real video but drops to $62.0$% on Omni clips. We introduce CamRM, which combines video frames with motion cues to judge camera-motion alignment. We develop CamRM on both the proprietary Gemini model and the open-source Qwen model, with consistent improvements of over $10$% across both. We also develop a training-free agentic model (CamRM-A), showing that our approach requires no fine-tuning. Our CamRM achieves $80.7$% accuracy and improves balanced accuracy over the strongest baseline by $6.6$%. In the motion-guided refinement loop, using CamRM as RM improves human-judged alignment by $13.2$% for Veo 3.1 and $18.4$% for Omni Flash at $2.2\times$ the generation cost, reaching $83.7$% alignment in the Omni Flash T2V case. At $8$ generations, the loop reaches $88.3$% pooled candidate-set oracle alignment versus $78.4$% for i.i.d. resampling, and matches its $8$-generation oracle rate with three generations, without generator training.