TrajLoc is a research paper on controlling the motion of several objects at once in image-to-video generation, accepted to NeurIPS 2026. Instead of mixing all trajectories into one shared conditioning signal, it constrains each object separately inside the attention layers.
Method: Each object token's cross-attention weights are replaced with a Gaussian heatmap centered on that object's target location in every frame.
Identity and depth: Trajectory and depth share a learned per-object token interface, and identity is kept by encoding first-frame appearance in place of the object token.
Results: The authors report tests across six datasets with up to 20 controlled objects, and gains of +4.3 dB PSNR and a 51% lower trajectory end point error versus the strongest baselines, on CogVideoX 5B and WaN 2.1 14B.
Availability: The paper links to a project page, code, and checkpoints.
Sign in to comment.
TL;DR of the paper, generated automatically.
TrajLoc is a research paper on controlling the motion of several objects at once in image-to-video generation, accepted to NeurIPS 2026. Instead of mixing all trajectories into one shared conditioning signal, it constrains each object separately inside the attention layers.