0
0The paper introduce trajectory replay, an innovative framework designed to update continuous-time generative models with sequentially arriving data without suffering from catastrophic forgetting. By replacing expensive sample-level generative replay with velocity-field regression along the historical model's own trajectories, the method halves the expected simulation cost and enhances robustness against discretization errors. Mathematically, the authors demonstrate that the optimum of this approach serves as a density-weighted mixture of historical and incoming probability paths, which effectively recovers a weighted forward KL divergence. Furthermore, the integration of a matching rule for the mean velocity enables efficient few-step sampling across the flow map. Extensive evaluations validate that trajectory replay outperforms traditional replay and distillation baselines on class-incremental image generation tasks. Finally, empirical assessments on biological case studies confirm its efficacy in extending generative perturbation models to new experimental contexts and modalities while successfully preserving treatment-effect predictions.
References:
Palma A, Jones C, Cinquin T, et al. Trajectory Replay: Continuous-Time Mixture Transport for Lifelong Generative Models[J]. arXiv preprint arXiv:2610.26749, 2026.
