A diffusion model trained on CIFAR-10 re-images black hole M87* — and the ring persists

bravo_abad · x · 2026-11-16

The Event Horizon Telescope's 2019 image of M87 has long faced a question: does the bright ring come from data or from imaging priors? Lai and colleagues re-imaged M87 with a diffusion model whose training set was 89% CIFAR-10—frogs, trucks, ships—and which had never seen an astronomical image. If a ring still appears, it came from the data. Context: EHT samples the sky's Fourier transform so sparsely that many images fit, and the prior decides the result; some reanalyses found no clear ring, and a standard pipeline initializes from a 40-microarcsecond Gaussian, about the size of the expected answer.

Original post →

More from Research

Research channel →