TRACE: New conformal prediction for diffusion and flow models

Constructing valid and informative conformal prediction regions for multi-dimensional outputs remains a fundamental challenge in machine learning. While conformal prediction offers finite-sample, distribution-free coverage guarantees, its practical performance hinges on the choice of nonconformity score. Existing methods often rely on restrictive geometric assumptions or require explicit likelihood evaluation and invertible transformations, limiting their use with complex generative models like diffusion and flow matching.

TRACE addresses this by defining nonconformity through transport alignment. Instead of evaluating likelihoods, it measures how well a candidate output aligns with the learned generative dynamics—averaging denoising or velocity-matching errors along stochastic transport trajectories from the generative process. The resulting scalar scores are calibrated using split conformal prediction, yielding valid marginal coverage under exchangeability. The framework is robust to computational budget, and experiments demonstrate valid coverage on synthetic and real datasets, with prediction regions that adapt naturally to multimodal and non-convex conditional distributions. TRACE opens the door to reliable uncertainty quantification for modern generative models without restrictive assumptions.