JEPA-Guided Diffusion for World-Centric Minority Sampling
A recent paper on arXiv (2605.24631) presents a world-centric method for minority sampling, focusing on creating low-density instances within a data manifold for uses such as medical diagnosis, anomaly detection, and creative AI. Current techniques determine minority samples based on generative priors from training datasets, which may not accurately capture real-world semantics. The authors propose JEPA guidance, a diffusion sampling framework influenced by a Joint-Embedding Predictive Architecture (JEPA), a type of world model that encapsulates semantically rich representations. This guidance directs diffusion paths toward low-density areas shaped by the implicit density from the JEPA, ensuring that generated minority samples align with real-world semantic contexts.
Key facts
- Paper arXiv:2605.24631 proposes JEPA-guided diffusion for minority sampling.
- Minority sampling generates low-density instances on a data manifold.
- Applications include medical diagnosis, anomaly detection, and creative AI.
- Existing methods define rarity relative to generative priors from training data.
- Proposed approach defines rarity with respect to real-world priors.
- JEPA stands for Joint-Embedding Predictive Architecture.
- JEPA guidance aligns generated minorities with real-world semantic priors.
- The paper was announced as a cross type on arXiv.
Entities
Institutions
- arXiv