Learning cosmic web environments with diffusion models

What does a diffusion model learn about the structure of the cosmic web? We train a model on Quijote simulations and compare its self-attention maps with voids, walls, filaments, and nodes identified by the T-Web classifier. Different layers capture overdense and underdense environments at different spatial scales, with cosmological information concentrated at intermediate and large scales. These results show that diffusion models learn a multi-scale representation of cosmic structure, including information beyond two-point statistics.