ConceptLangevin dynamics

Langevin Dynamics

A sampling method based on stochastic gradient descent that explores a distribution by injecting noise into the energy gradient. It converges to the target distribution in theory but is relatively slow.

Timeline (last 90 days)

Oct 26

Gaussian Splatting EBM supports two sampling methods, Rectified Flow and Langevin dynamics; in Phase-1 testing, Rectified Flow sampling was 29-39 times faster than 200-step Langevin.

Unverified50%

All Facts (1)

Unverified

Gaussian Splatting EBM supports two sampling methods, Rectified Flow and Langevin dynamics; in Phase-1 testing, Rectified Flow sampling was 29-39 times faster than 200-step Langevin.

50%

Source Articles

Averaged Annealed Markov Chains Achieve the Algorithmic Threshold for Tensor PCAOct 30 Gaussian-Splat EBM: AUROC on Par with kNN Density Estimation, 3.9x Faster InferenceOct 26