CamGeo: Sparse Camera-Conditioned Image-to-Video Generation with 3D Geometry Priors

Sparse camera-conditioned image-to-video generation presents a pivotal challenge: synthesizing geometrically consistent 3D motion from minimal pose cues. Existing methods, which largely rely on dense supervision or naive interpolation, suffer from severe pose drift and motion discontinuities due to the lack of robust 3D priors. In this paper, we introduce \textbf{CamGeo}, a novel framework that distills rich 3D geometric knowledge from a pre-trained video-to-3D model (VGGT) directly into the diffusion backbone. To achieve this without incurring inference latency, we propose a training-only distillation strategy. Specifically, CamGeo incorporates: (1) keyframe trajectory distillation that enforces cycle-consistency with sparse input poses, (2) cross-frame consistency distillation with both camera trajectory and depth constraints to generate consistent structure across unsupervised frames, and (3) a three-stage coarse-to-fine curriculum learning, progressively scales geometric complexity, from global structure coherence to fine-grained refinement, achieving stable optimization. Extensive experiments demonstrate that CamGeo achieves consistent improvements under various sparsity ratios.

Many image-to-video systems can animate a single picture, but they are still difficult to control when users want a specific camera movement. Existing camera-controlled methods often rely on dense camera trajectories, which are hard to obtain in everyday use. In practice, a user may only know a few key viewpoints, such as where the camera should start, turn, or end. We introduce CamGeo, a method that allows image-to-video models to generate videos from sparse camera conditions. During training, CamGeo uses a frozen 3D vision model to provide additional guidance about camera motion and scene geometry. This helps the video model learn how to fill in plausible motion between the few provided camera poses. At test time, the 3D model is no longer needed, so the method does not add extra inference cost. This makes camera-controllable video generation more practical: users can specify fewer camera inputs while still obtaining videos with coherent viewpoint changes and stable geometry. The approach can support applications such as virtual scene creation, cinematic camera design, and immersive content generation.