Light from distant galaxies travels to us for billions of years, but along the way it gets bent and twisted by the invisible web of dark matter. This effect, predicted by Albert Einstein within general relativity, is called gravitational lensing. Back in the 1930s, Fritz Zwicky suspected hidden mass in clusters, and later Vera Rubin proved its ubiquity through the frantic rotation of galaxies. Today we collect colossal data about the coherent distortions of shapes of billions of galaxies: this is weak lensing, and it turns the entire sky into one sensitive detector of the cosmic web.

Traditionally, only averaged correlations were extracted from these data, but that's like trying to reconstruct a person's face from a blurry silhouette on fogged glass — fine features, wrinkles, and moles that carry the main information are lost. Non-Gaussian patterns, vortices and clumps born from gravitational collapse, hold the keys to the mystery of accelerated expansion of the Universe and the nature of dark energy. The JADE method works differently. It immediately tackles the full picture: by training on hundreds of thousands of numerical simulations of cosmic fields, a diffusion model learns to reverse the distortions, step by step peeling away the gravitational veil. At the heart of the architecture lies a transformer that intertwines the pixels of a map with six cosmological parameters, like a neural polyglot fluent in the language of images and the algebra of constants. The result is not one map, but an entire gallery of probable universes, where each sample gives birth to a consistent realization of the mass field.

Validation on simulated data from the LSST Y10 survey showed brilliant results: the reconstructed tomographic slices of redshift not only accurately reproduce the power spectrum (with an error of about 7%), but also, as if tracing the imprint of an invisible palm, precisely capture the one-point distribution — a direct measure of non-Gaussianity. The cosmological parameters, including dark matter density Ωc, fluctuation amplitude σ8, and the equation of state of dark energy w0, converged with classical MCMC chains as if a riddle solved by different detectives pointed to the same suspect. The key breakthrough is the abandonment of a differentiable simulator. Previously, physics had to be simplified into differentiable toys; now JADE accepts heavyweight N-body and hydrodynamical codes as they are, without sacrificing honest Bayesian inference. It's like allowing a detective to finally use the full arsenal of clues, not just simple fingerprints.

The prospects of JADE extend far beyond flat maps. The next step is adaptation to the full celestial sphere and inclusion of gradients from direct simulations once they become available. The architecture is already eyeing multi-probe analysis, ready to weave together lensing, galaxy clusters, and the cosmic microwave background, and even to recover initial conditions — the quantum seeds from which the cosmic web grew. After all, up to half of all information about dark energy is hidden in non-Gaussian patterns, and JADE, for the first time, extracts it without compromise. This brings us closer to the dream of cosmologists: not just to measure parameters, but to peer into the hidden geometry of the cosmos through the distorted light of distant stars, as if watching a movie about how the invisible became visible.