CARV Framework Reduces Variance in Diffusion Teacher Pipelines

A new framework called CARV (Compute-Aware Variance Accounting) aims to reduce estimator variance in pipelines that use pretrained diffusion models as frozen teachers, such as text-to-3D generation, single-step distillation, and data attribution. These pipelines rely on Monte Carlo expectations over noise levels and Gaussian noise samples, where each draw requires expensive upstream computations like rendering, simulation, or encoding. CARV introduces a hierarchical Monte Carlo estimator that amortizes expensive upstream computation over cheap diffusion-noise resamples, enhanced by timestep importance sampling and a stratified-inverse-CDF construction. Experiments in text-to-3D distillation and attribution show CARV delivers 2-3x effective compute multipliers, with most gains from amortized reuse and about 25% additional from importance sampling and stratification. In single-step distillation, the same techniques cut gradient variance significantly. The framework does not change the objective function. The paper is available on arXiv under ID 2605.21489.