FluIDWorld swaps attention for PDE dynamics

The interesting bit here is not that a PDE can compete on one-step metrics; it’s that the inductive bias seems to pay off once the model has to survive its own predictions.

• –The comparison is unusually fair: PDE, Transformer, and ConvLSTM are all kept near 800K parameters with the same encoder, decoder, losses, and data.
• –The paper’s main win is rollout stability, where diffusion behaves like an implicit spatial regularizer and slows error accumulation.
• –Single-step parity plus better long-horizon coherence is exactly the kind of result that could make world-model research less Transformer-centric.
• –The O(N) local-update story is compelling for efficiency-minded teams, especially since the experiments were run on a single consumer GPU.