Research map / Image generation / Text-to-image diffusion
Your Diffusion Model is Secretly a Zero-Shot Classifier
Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, Deepak Pathak
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the diffusion model's noise-prediction error as a zero-shot classifier instead of training a ResNet-50 on synthetic SD images | Stable Diffusion 2.0 · ImageNet (zero-shot) | accuracy: 18.9 -> 61.4 | Improved |
| use the same (t, eps) samples for all classes (difference testing) (ablation) | Diffusion Classifier · Pets | estimate accuracy | Improved |
| sample timesteps evenly across [1,1000] instead of repeating a few intermediate timesteps (ablation) | Diffusion Classifier · Pets | accuracy scaling | Improved |
| use uniform l2 weighting instead of variational lower bound as objective (ablation) | DiT-XL/2 Diffusion Classifier at 512^2 · ImageNet | accuracy: 74.0 -> 79.1 | Improved |
| use l1 error instead of squared l2 in zero-shot classifier (ablation) | Stable Diffusion 2.0 · Zero-shot datasets (8) | accuracy: ImageNet 61.4 -> 59.6; CIFAR10 84.4 -> 88.4 | Mixed |
| use l1 error instead of squared l2 in DiT supervised classifier (ablation) | DiT-XL/2 at 512^2 · ImageNet-A | accuracy: 30.2 -> 13.2 | Hurt |
| prune classes with CLIP ResNet-50 top-20 before diffusion scoring (ablation) | Diffusion Classifier · ImageNet (zero-shot) | accuracy / time per image: 61.4 -> 62.6 accuracy; 1000s -> 150s | Improved |
| upgrade Stable Diffusion from 2.0 to 2.1 (ablation) | Diffusion Classifier · ImageNet (zero-shot) | accuracy: 61.4 -> 58.4 | Hurt |
| use later SD 1.x checkpoints with more training (ablation) | Diffusion Classifier · ImageNet (zero-shot) | accuracy: 51.7 -> 59.6 (SD 1.1 to 1.5) | Improved |
| add classifier-free guidance to the noise-prediction error (ablation) | Diffusion Classifier · zero-shot classification | accuracy | Hurt |
| classify with generative Diffusion Classifier instead of discriminative classifier | DiT-XL/2 vs ResNet/ViT · ImageNet-A | accuracy: 20.8 -> 30.2 (ViT-B/16 vs 512^2) | Improved |
| generative classifier effective robustness on ImageNetV2 and ObjectNet | DiT-XL/2 Diffusion Classifier · ImageNetV2, ObjectNet | effective robustness | No clear effect |
Closest papers by meaning
- Diffusion Classifiers Understand Compositionality, but Conditions Apply (NeurIPS 2025 · similarity 0.94)
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.89)
- Mining your own secrets: Diffusion Classifier Scores for Continual Personalization of Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.89)
- Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Diffusion Self-Distillation for Zero-Shot Customized Image Generation (CVPR 2025 · similarity 0.89)
- Dual Diffusion for Unified Image Generation and Understanding (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
