Shaping Useful Noise: Energy Distributions Predict Visual Pretraining Quality

Synthetic data is useful when it occupies the right real-image likelihood regimes; ELD measures energy shape to predict, diagnose, curate visual pretraining data.

Energy-based data auditing

Energy-Level Distribution Distance compares the energy distributions of real and synthetic images under the same normalised real-image density model. The interactive project page lets you explore illustrative energy profiles and the reported DINO/ViT-S results.