Mint

[NeurIPS 2025] Mint: A Simple Test-Time Adaptation of Vision-Language Models against Common Corruptions. Wenxuan Bao*, Ruxi Deng*, Jingrui He

Key packages

torch          2.4.1
torchvision    0.19.1
clip           1.0

Datasets

1. Downloads data from the corruption benchmark (CIFAR-10-C, CIFAR-100-C, and ImageNet-C) and extract them.
2. Download classnames.txt from this link.
3. The directory structure should look like

${data_root}/corruption
│ 
├── CIFAR-10-C
│   ├── brightness.npy
│   ├── ...
│   ├── pixelate.npy
│   └── labels.npy
│ 
├── CIFAR-100-C
│   ├── brightness.npy
│   ├── ...
│   ├── pixelate.npy
│   └── labels.npy
│ 
└── CIFAR-100-C
    ├── brightness
    │   ├── 1
    │   ├── ...
    │   └── 5
    │       ├── n01440764
    │       │   ├── ILSVRC2012_val_00000293.JPEG
    │       │   ├── ...
    │       │   └── ILSVRC2012_val_00048969.JPEG
    │       ├── ...
    │       └── n07579787
    ├── ...
    ├── pixelate
    └── classnames.txt

Experiments

To run experiments:

cd ./shell
./run_mint.sh