Research & Innovation
OREO Improves a 3D Generator by Letting a 2D Model Critique Its Renders
3D generators still produce assets that look subtly wrong. OREO, posted September 24, skips static datasets: it renders the 3D output, has a 2D diffusion model refine those views while preserving geometry, viewpoint and content, then uses the refined views as supervision — so the generator learns from improved versions of its own samples.