---
license: mit
tags:
  - video-deblurring
  - diffusion
  - replica
---

# VD-Diff trained on Replica

Checkpoints from [VD-Diff](https://github.com/Chen-Rao/VD-Diff) (Rethinking Video Deblurring with Wavelet-Aware Dynamic Transformer and Diffusion Model, ECCV 2024) trained on the Replica scenes.

Code that produced these checkpoints: <https://github.com/1zsw123/pretraineddatasetcode>

## Files

| File | Stage | Step | Size |
|---|---|---|---|
| `S1/net_g_200000.pth` | Stage 1 (KPN pre-train) | 200k | 38 MB |
| `S3/net_g_175000.pth` | Stage 3 (diffusion fine-tune) | 175k | 48 MB |

## How to use

```python
from huggingface_hub import hf_hub_download
s1 = hf_hub_download("zhaoshiwen/VD-Diff-Replica", "S1/net_g_200000.pth")
s3 = hf_hub_download("zhaoshiwen/VD-Diff-Replica", "S3/net_g_175000.pth")
```

Then point `pretrain_network_S1` / `pretrain_network_S2` in the VD-Diff training YAML at these paths. See the upstream README for the training pipeline.