Accelerating black hole image generation via latent space diffusion models
Phys. Rev. D 114, 083049 – Published 27 October, 2026
DOI: https://doi.org/10.1103/78xb-q47l
Abstract
Interpreting horizon-scale black hole images currently relies on computationally intensive general relativistic ray-tracing simulations, which pose a significant bottleneck for rapid parameter exploration and high-precision tests of strong-field gravity. We demonstrate that physically accurate black hole images, synthesized from magnetized accretion flows, inherently reside on a low-dimensional manifold encoding the essential features of spacetime geometry, plasma distribution, and relativistic emission. Leveraging this structure, we introduce a physics-conditioned diffusion model that operates in a compact latent space to generate high-fidelity black hole imagery directly from physical parameters. The model reproduces key general relativistic ray-tracing signatures, including shadow diameter, photon-ring structure, and relativistic brightness asymmetry, while reducing the average inference time from 5.25 to 1.15 s per image relative to the baseline model. Ablation studies show that self-attention drives the improvements in reconstruction and parameter estimation, whereas principal component analysis provides computational efficiency through latent compression.
