Get Started
Research questionHow can training-data clusters be attributed to flow-matching images when influence propagates through the generation trajectory?A training sample can change the velocity field at one point, while the effect of that change evolves before reaching the final image. Consequently, local field changes may not predict a sample’s eventual counterfactual influence.
Image Generation
Machine Learning
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.Tracing Generated Samples to Training-Data Clusters in Flow-Matching ModelsThe evidence covers trajectory-based attribution in two flow-matching latent spaces, compared with semantic-similarity and other baselines using independently retrained leave-one-cluster-out models. The reported approach includes a closed-form score that does not require counterfactual retraining or model gradients, but its competitiveness varies across evaluation metrics.research paper · Sep 2, 2026
Related questions
How can pixel-space flow-matching objectives avoid low-frequency dominance while learning fine image details?How can Flow Matching avoid crossing paths and velocity ambiguity without costly optimal-transport coupling?How can learned models approximate trajectory similarity across multiple granularities without losing relative similarity rankings?How can training-data attribution distinguish reweighting influence from genuine intervention leverage?
Home
Topics
Search
Library