SymDrift: One-Shot Generative Modeling under Symmetries
SymDrift is a newly introduced framework that tackles a specific challenge related to symmetry in the generative modeling of physical systems, including molecules. While equivariant diffusion and flow matching models can effectively integrate global symmetries like 3D rotations, they often involve expensive multi-step sampling processes. In contrast, drifting models allow for efficient single-step generation but struggle because an equivariant generator typically fails to yield the same drifting field as the symmetrized target distribution, leading to costly symmetrization. By aligning the drifting process with symmetries, SymDrift eliminates this expense, facilitating one-shot generation while maintaining invariance. Further details can be found in arXiv:2605.06140.
Key facts
- SymDrift is a framework for generative modeling under symmetries.
- It addresses a challenge specific to drifting models.
- Equivariant diffusion and flow matching models are costly due to multi-step sampling.
- Drifting models enable single-step generation.
- An equivariant generator does not generally match the symmetrized target distribution.
- SymDrift avoids expensive symmetrization of the empirical distribution.
- The work is published on arXiv with ID 2605.06140.
- The method targets physical systems such as molecules.
Entities
Institutions
- arXiv