Research map / Label-efficient and robust learning / Test-time adaptation
Diffusion-calibrated Continual Test-time Adaptation
Xu Yang, Moqi Li, Kun Wei
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 |
|---|---|---|---|
| add generated-instance cross-entropy loss (CE1) (ablation) | CE baseline · ImageNet-to-ImageNet-C (ViT-base) | mean error rate (%): 51.5 -> 46.8 | Improved |
| add reliable-instance selection and diversity weighting (WCE) (ablation) | CE + CE1 · ImageNet-to-ImageNet-C (ViT-base) | mean error rate (%): 46.8 -> 43.2 | Improved |
| add contrastive learning with diffusion negatives (ablation) | WCE · ImageNet-to-ImageNet-C (ViT-base) | mean error rate (%): 43.2 -> 38.3 | Improved |
| contrastive learning over all instances X instead of the full CL (ablation) | WCE · ImageNet-to-ImageNet-C (ViT-base) | mean error rate (%): 41.8 vs 38.3 | Hurt |
| contrastive learning only on unselected instances (X \ S) (ablation) | WCE · ImageNet-to-ImageNet-C (ViT-base) | mean error rate (%): 42.5 vs 38.3 | Hurt |
| contrastive learning only on selected instances S (ablation) | WCE · ImageNet-to-ImageNet-C (ViT-base) | mean error rate (%): 41.2 vs 38.3 | Hurt |
| diffusion-calibrated supervision versus ROID | ROID · ImageNet-to-ImageNet-C (ViT-base) | mean error rate (%): 45.0 -> 38.3 | Improved |
| diffusion-calibrated supervision versus REM | REM · ImageNet-to-ImageNet-C (ViT-base) | mean error rate (%): 39.2 -> 38.3 | Improved |
| diffusion-calibrated supervision versus ROID | ROID · ImageNet-to-ImageNet-C (ResNet) | mean error rate (%): 54.5 -> 50.2 | Improved |
| diffusion-calibrated supervision versus REM | REM · CIFAR100-to-CIFAR100C (ViT-base) | mean error rate (%): 23.4 -> 22.4 | Improved |
| diffusion-calibrated supervision versus REM | REM · CIFAR10-to-CIFAR10C (ViT-base) | mean error rate (%): 9.4 -> 8.7 | Improved |
| diffusion-calibrated supervision versus ROID | ROID · CIFAR100-to-CIFAR100C (ResNet) | mean error rate (%): 29.3 -> 26.7 | Improved |
Closest papers by meaning
- A Versatile Framework for Continual Test-Time Domain Adaptation: Balancing Discriminability and Generalizability (CVPR 2024 · similarity 0.93)
- Everything to the Synthetic: Diffusion-driven Test-time Adaptation via Synthetic-Domain Alignment (CVPR 2025 · similarity 0.91)
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.88)
- ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation (ICLR 2024 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.87)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.87)
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.
