Realistic subsurface models are hard to condition and tractable statistical ones lack geological realism. We use a Vision Transformer as an approximate prior with sampling corrections for posterior inference and a built-in proposal-quality diagnostic.