Research map / Image generation / Fast diffusion sampling

Discriminative Score Function: Turning Pretrained Models into Functional Generative Priors

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Distributional control of diffusion

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces the Discriminative Score Function (DSF), a training-free image-space generative field extracted from the loss-gradient geometry of pretrained discriminative models (ResNet, DETR, DINOv2), used for generation, editing, inpainting, explanation and zero-shot quantization calibration.

Problem

Discriminative vision models are trained for task objectives, yet may retain implicit data-manifold structure; it is unclear whether this can be exposed as a label-free generative signal for synthesis and calibration across diverse architectures.

Approach

Uses the Loss Tangent Kernel as a training-free functional proxy: image updates align the loss-gradient of a label-free augmentation-invariance loss with the pretrained parameter direction (-θ as a proxy for the aggregate training gradient), yielding a score-like update that moves noise images toward model-consistent regions.

Key results

• On zero-shot SmoothQuant quantization of DINOv2, DSF calibration at 4-bit reaches k-NN .589 and CLS .710, versus MimiQ .078 k-NN and Gaussian .077 k-NN in the same setting (real-image oracle .580).
• DSF matches or exceeds the real-image oracle in nearly all 8/6/4-bit settings for feature-geometry metrics (e.g., 4-bit patch similarity .760 vs .753 for real images).
• DSF generates unconditional images from noise using ResNet, DETR, and DINOv2 without labels, real data, or external generators, at about 1.4-1.7 s per image on an A6000.

Surprising findings

• Generated samples from ResNet repeatedly show dataset-typical objects such as analog clocks, exposing dataset biases.
• DSF calibration slightly exceeds the real-image oracle at 4-bit on several feature-geometry metrics.

Limitations

• Not a replacement for high-fidelity diffusion generators; sample quality is limited.
• Gradient proxy depends on architecture and training recipe.
• Parameter-gradient (second-order) updates make sampling costlier than first-order image optimization.

Experimental setup

Datasets: ImageNet COCO LVD-124M

Benchmarks: ImageNet

Metrics: k-NN top-1 linear head accuracy fpQ@5 CLS cosine similarity patch cosine similarity

Baselines: RegCache Gaussian noise calibration MimiQ real ImageNet calibration images feature visualization (Ghiasi et al.)

Builds on: Loss Tangent Kernel Neural Tangent Kernel classifier-guided diffusion SmoothQuant

Closest papers by meaning

• Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator (ICML 2025 · similarity 0.87)
• Intriguing Properties of Generative Classifiers (ICLR 2024 · similarity 0.87)
• An Efficient Framework for Enhancing Discriminative Models via Diffusion Techniques (AAAI 2025 · similarity 0.86)
• Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.86)
• Emphasizing Discriminative Features for Dataset Distillation in Complex Scenarios (CVPR 2025 · similarity 0.86)
• UMFC: Unsupervised Multi-Domain Feature Calibration for Vision-Language Models (NeurIPS 2024 · similarity 0.86)