A 2D point cloud that morphs in real time as a stack of invertible coupling-layer transforms warps a simple Gaussian base distribution into a complex target data distribution, with a live readout of how the Jacobian determinant rescales density at each step.
Choose a target data distribution, add or remove coupling layers with the depth slider, and press play to watch the Gaussian cloud flow layer by layer into the target shape while the log-likelihood readout updates in real time.
Target distribution select, coupling-layer depth slider, play/pause flow animation, log-likelihood readout toggle, rebuild flow
Because every flow layer is exactly invertible, you can run a trained normalizing flow backward on a real data point to recover its latent code precisely, then forward again to reconstruct the exact original sample with zero reconstruction error.
A 2D point cloud that morphs in real time as a stack of invertible coupling-layer transforms warps a simple Gaussian base distribution into a complex target data distribution, with a live readout of how the Jacobian determinant rescales density at each step.
A 2D point cloud that morphs in real time as a stack of invertible coupling-layer transforms warps a simple Gaussian base distribution into a complex target data distribution, with a live readout of how the Jacobian determinant rescales density at each step.
Choose a target data distribution, add or remove coupling layers with the depth slider, and press play to watch the Gaussian cloud flow layer by layer into the target shape while the log-likelihood readout updates in real time.
Because every flow layer is exactly invertible, you can run a trained normalizing flow backward on a real data point to recover its latent code precisely, then forward again to reconstruct the exact original sample with zero reconstruction error.