MagGN · magnitude distance · interactive figure
X stands in for a batch of real data and Y for a batch from a generator. Drag Y across the plane, change the scale t or break the generator, and the normalized magnitude distance that MagGN minimises is recomputed in your browser with the same formula as magnitude.py.
Real and generated batches
Drag anywhere to move Y. Point size shows each point's weight in |X ∪ Y|t.
Distance across scales
The same two batches, every t from 0.1 to 100
Distance as Y slides away
Y translated along the current shift direction, one curve per t
At this scale
|·|t = sum of the entries of S−1, the effective number of distinct points
Magnitude counts how many distinct points a set has at scale t. If Y lands on top of X, the union adds almost nothing, so |X ∪ Y| ≈ |X| and the distance falls to the sampling floor. If Y sits far away, the union counts both, |X ∪ Y| ≈ |X| + |Y|, and the normalized distance reaches 1.
Small t sees coarse layout and saturates once the batches are apart. Large t sees individual modes but is near 1 for any two different samples. The toy MagGN run averages t = 5, 20 and 50 so the generator gets a gradient at every stage. Try mode dropping on 8 Gaussians to see which scales notice.
The page builds S = exp(−t·D) for X, Y and X ∪ Y, factors each with a float64 Cholesky and sums S⁻¹1. Training uses eps = 0; when a matrix is too close to singular (tight clusters at small t) the page adds the smallest jitter that works and says so.
Distance: norm_diff_magnitude_distance_grad(X, Y, t, normalize=True) = (2|X∪Y|t − |X|t − |Y|t) / |X∪Y|t. Datasets follow Toy_Experiment/data_generator.py; multiscale schedule from MagGN_toy_experiments.ipynb. Points are sampled in the browser with a fixed seed, not loaded from training runs. From “Magnitude Distance: A Geometric Measure of Dataset Similarity”, ICML 2026.