The cosmos resembles silk fabric with threads of the dark web showing through. We only glimpse tiny swatches of the real tapestry — through telescopes, in actual galaxy catalogs. To separate the fundamental signal from astrophysical noise in the cosmic pattern, statistical analysis demands thousands of plausible sketches — synthetic sky maps. But how to create them? For a long time, astronomers acted like patient weavers: they ran full-blown numerical simulations (virtual worlds with gas, stars, and black holes) consuming months of supercomputer time.

A different technique is automatic embroidery. The role of the skilled artisan is played by a diffusion model (a neural network that learns to remove noise from random data and recover hidden patterns). It is given only a faint sketch — a low-resolution 3D cube of dark matter density, deliberately blurred, like a drawing on tracing paper with smudged lines. From this sketch, the embroiderer-network returns a detailed picture in seconds: the precise positions of galaxies, their stellar masses, and the rates at which new stars are born. The backbone of the architecture is a transformers with a convolutional Vision Transformer for encoding the density field, plus attention mechanisms: each galaxy-point 'looks' at all the others, refining the overall ensemble.

Trained on 90 000 such swatches from the IllustrisTNG simulation, after lengthy training the model produces catalogs that even an expert cannot distinguish from the ground truth. Galaxies settle onto the invisible filaments and nodes of the dark web, and histograms of star formation rates and power spectra of the large-scale structure match the true values within statistical errors. Most striking of all, the embroiderer guesses the galaxy population in small dark matter clumps that are unresolved in the original rough draft: it marginalizes over sub-grid fluctuations, essentially reconstructing what it never 'saw'. The entire cube, 151.3 megaparsecs on a side, is filled in 10 seconds on a GPU.

This approach is not just an elegant trick: it changes the technology of analysis. Instead of a long chain of 'simulation → halo finding → galaxy occupation', we get an end-to-end generator that directly translates a density field into observed objects. This reduces systematic errors and simplifies the Machine Learning Bayesian inference of cosmological parameters. By adding a global parameter to the conditioning, redshift or a different feedback recipe, one obtains a generator for different epochs and physical scenarios. In this way, small-scale clustering, where the lion's share of the signal hides, becomes accessible for extraction. The dark canvas no longer seems mute: the rapid stitches of neural networks reveal on it a map of the Universe that once required years of manual work.