Making diffusion model gradients 3× cheaper to compute

Diffusion models train text-to-3D systems and other pipelines, but computing gradients through them is expensive—each gradient requires costly upstream work like rendering or simulation. CARV reuses that expensive computation across multiple noise samples, then sharpens estimates with importance sampling and stratification. The method delivers 2–3× speedups in real tasks like 3D distillation and data attribution, though single-step distillation shows the technique hits diminishing returns when other bottlenecks dominate.