Research map / Image generation / Fast diffusion sampling
Diffusion-TTA: Test-time Adaptation of Discriminative Models via Generative Feedback
Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, Katerina Fragkiadaki
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 a single random timestep and noise latent for diffusion loss TTA (ablation) | Diffusion-TTA (ConvNext-Tiny) · ImageNet-A | accuracy: 22.7 -> 19.8 | Hurt |
| add timestep augmentation with batch size 180 (ablation) | Diffusion-TTA (ConvNext-Tiny) · ImageNet-A | accuracy: 19.8 -> 24.5 | Improved |
| add random noise latents per timestep (ablation) | Diffusion-TTA (ConvNext-Tiny) · ImageNet-A | accuracy: 24.5 -> 24.7 | Improved |
| also adapt the diffusion model weights (ablation) | Diffusion-TTA (ConvNext-Tiny) · ImageNet-A | accuracy: 24.7 -> 25.8 | Improved |
| randomly initialize the classifier instead of pretrained weights (ablation) | Diffusion-TTA (ResNet18) · ImageNet | accuracy: 68.4 -> 0.7 | Hurt |
| adapt only logits with diffusion objective, no classifier (ablation) | Diffusion-TTA (ResNet18) · ImageNet-R | accuracy: 37.0 -> 31.0 | Hurt |
| adapt only BN layers (ablation) | Diffusion-TTA (ResNet18) · ImageNet | accuracy: 68.4 -> 69.0 | No clear effect |
| adapt only last FC layer (ablation) | Diffusion-TTA (ResNet18) · ImageNet | accuracy: 68.4 -> 68.6 | No clear effect |
| also adapt the diffusion model (DiT) in addition to the classifier (ablation) | Diffusion-TTA (ResNet18) · ImageNet | accuracy: 73.0 -> 78.2 | Improved |
| apply Diffusion-TTA to CLIP-ViT-B/16 | CLIP-ViT-B/16 · CIFAR-100 | accuracy: 69.4 -> 69.0 | Hurt |
| apply Diffusion-TTA to CLIP-ViT-L/14 | CLIP-ViT-L/14 · Food101 | accuracy: 93.1 -> 93.1 | No clear effect |
| Diffusion-TTA vs TENT online on ConvNext-Tiny | ConvNext-Tiny · ImageNet-A | accuracy: 10.6 -> 25.8 | Improved |
Closest papers by meaning
- Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment (CVPR 2025 · similarity 0.90)
- Exploring Structured Semantic Priors Underlying Diffusion Score for Test-time Adaptation (NeurIPS 2024 · similarity 0.90)
- Not All Diffusion Model Activations Have Been Evaluated as Discriminative Features (NeurIPS 2024 · similarity 0.89)
- Diffusion Classifiers Understand Compositionality, but Conditions Apply (NeurIPS 2025 · similarity 0.89)
- AP-Adapter: Improving Generalization of Automatic Prompts on Unseen Text-to-Image Diffusion Models (NeurIPS 2024 · similarity 0.89)
- An Efficient Framework for Enhancing Discriminative Models via Diffusion Techniques (AAAI 2025 · similarity 0.88)
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.
