New TrajLoom framework predicts future motion in videos by modeling dense point trajectories using flow matching and variational autoencoders.

TrajLoom uses Grid-Anchor Offset Encoding to reduce location bias by representing points as offsets from pixel-center anchors

TrajLoom-VAE learns compact spatiotemporal latent space for dense trajectories with masked reconstruction and consistency regularization

TrajLoom-Flow generates future trajectories via flow matching with boundary cues and K-step fine-tuning for stable sampling

Framework predicts both future trajectories and visibility from observed video context for improved motion representation

Introduces TrajLoomBench, a unified benchmark dataset combining real and synthetic video data for evaluation