Cross Attention News

TrajLoc tackles a nasty multi-object video-control problem by replacing dense shared conditioning with per-object attention masks. In practice, that means better identity preservation when trajectories cross or objects occlude each …

For animation teams, the useful bit is that MultiAct can steer pretrained text-to-motion models toward composite prompts without retraining them. It tackles the common failure mode where a model latches …

A training-free diffusion technique can now suppress unwanted concepts by steering cross-attention only when the target appears, instead of retraining the whole model. For teams using generative tools, that means …