Image-to-Video Generation

Image-to-video generation studies how to turn a still image into a moving clip while preserving the subject, scene, and visual style. Most systems extend pretrained video diffusion models, using temporal attention, lightweight adapters, or controls such as poses, trajectories, masks, depth, and camera paths; some add 3D geometry or physics simulation to guide motion. Training-free prompt refinement and inference-time search offer alternative ways to improve control, while distillation and sparse attention target faster generation. Evaluation increasingly tests motion adherence, identity and scene consistency, physical plausibility, and safety, but realistic appearance and smooth motion still do not guarantee correct physics or precise control.