DyPSI is a dynamic physics‑sensing framework that formulates the generation of future fields and sensor trajectories as a joint task conditioned on historical observations. Discrete sensor coordinates are lifted into continuous Sensor Position Fields, and both fields and sensing configurations are encoded with a shared Functional Tucker representation. A joint diffusion model is trained in the resulting latent space to generate future fields and trajectories simultaneously. To enforce temporal smoothness, the perturbation injected during diffusion training is correlated by a Gaussian‑process temporal kernel, biasing the denoiser toward smooth recoveries. The authors further relate the EDM training objective to a sequence‑aggregated A‑optimality criterion and to a kernel‑induced bound on discrete‑trajectory variation. Experiments on turbulent flow, GLORYS12 sea‑surface temperature, and 3‑D car aerodynamics demonstrate that DyPSI consistently reduces error under tight sensor budgets, outperforming static and frame‑wise placement strategies. The code is publicly available.\n\nReview: By continuous‑valued sensor positioning and temporally‑aware diffusion training, DyPSI achieves spatio‑temporal coherent sensing, highlighting the promise of generative models for physics‑field reconstruction.