---
license: other
license_name: research-only
tags:
  - anomaly-segmentation
  - out-of-distribution-detection
  - normalizing-flows
  - autonomous-driving
library_name: pytorch
---

# SAFe — trained conditional normalizing-flow checkpoints

Anonymous checkpoint release accompanying a paper submission under double-blind
review. This repository hosts the trained normalizing-flow weights; all
code, configs, and instructions are in the anonymized code repository linked
from the paper.

## Contents

```
checkpoints/
  flows/                                  # trained conditional normalizing flows
    convnext_l_res4_cityscapes/checkpoint_best.ckpt
    convnext_l_res4_issu/checkpoint_best.ckpt
    convnext_l_res5_cityscapes/checkpoint_best.ckpt
    convnext_l_res5_issu/checkpoint_best.ckpt
    vit_l_res5_cityscapes/checkpoint_best.ckpt
    vit_l_res5_issu/checkpoint_best.ckpt
  priors/                                 # per-class Gaussian priors
    convnext_l_res4_cityscapes.pth
    convnext_l_res4_issu.pth
    convnext_l_res5_cityscapes.pth
    convnext_l_res5_issu.pth
    vit_l_cityscapes.pth
    vit_l_issu.pth
```


| Checkpoint | Backbone | Feature level | Training data |
|---|---|---|---|
| `convnext_l_res4_cityscapes` | DINOv3 ConvNeXt-L | res4 | Cityscapes (train) |
| `convnext_l_res5_cityscapes` | DINOv3 ConvNeXt-L | res5 | Cityscapes (train) |
| `vit_l_res5_cityscapes`      | DINOv3 ViT-L     | last block | Cityscapes (train) |
| `convnext_l_res4_issu`       | DINOv3 ConvNeXt-L | res4 | ISSU-static (train) |
| `convnext_l_res5_issu`       | DINOv3 ConvNeXt-L | res5 | ISSU-static (train) |
| `vit_l_res5_issu`            | DINOv3 ViT-L     | last block | ISSU-static (train) |

## Usage

Clone the anonymized code repository, then download this repo's `checkpoints/`
directory to the repo root:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="<ANON-USER>/safe-checkpoints",
    repo_type="model",
    local_dir=".",
    allow_patterns=["checkpoints/**"],
)
```



## License

Released for the purpose of peer review and research reproducibility only.