Discriminative models—classifiers, detectors, encoders—are built to label or represent data, not to generate it. The authors of this paper argue that such models still carry an implicit trace of the data manifold they were trained on, and they propose a way to expose it.

Their Discriminative Score Function (DSF) extracts a generative field from the loss-gradient geometry of a pretrained model. This yields a training-free update rule that nudges images toward regions the model considers consistent, without retraining or task-specific targets. The same field supports unconditional generation, conditional guidance, editing, inpainting, and explanation.

The authors demonstrate DSF across ResNet classifiers, DETR detectors, and DINOv2 ViT encoders. They also show that DSF-generated samples serve as effective zero-shot calibration data for quantization, preserving feature geometry better than existing data-free baselines. The paper is accepted at NeurIPS 2026.