README.md

---
language:
- en
tags:
- pytorch_model_hub_mixin
- animation
- video-frame-interpolation
- uncertainty-estimation
license: mit
pipeline_tag: image-to-image
---

# 🤖 Multi‑Input ResShift Diffusion VFI

<div align="left" style="display: flex; flex-direction: row; gap: 15px">
  <a href='https://arxiv.org/pdf/2504.05402'><img src='https://img.shields.io/badge/arXiv-2405.17933-b31b1b.svg'></a> 
  <a href='https://github.com/VicFonch/Multi-Input-Resshift-Diffusion-VFI'><img src='https://img.shields.io/badge/Repo-Code-blue'></a> 
  <a href='https://colab.research.google.com/drive/1MGYycbNMW6Mxu5MUqw_RW_xxiVeHK5Aa#scrollTo=EKaYCioiP3tQ'><img src='https://img.shields.io/badge/Colab-Demo-Green'></a>
  <a href='https://huggingface.co/spaces/vfontech/Multi-Input-Res-Diffusion-VFI'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face%20Space-Demo-g'></a>
</div>

## ⚙️ Setup

Start by downloading the source code directly from GitHub.

```bash
git clone https://github.com/VicFonch/Multi-Input-Resshift-Diffusion-VFI.git
```

Create a conda environment and install all the requirements

```bash
conda create -n multi-input-resshift python=3.12
conda activate multi-input-resshift
pip install -r requirements.txt
```

**Note**: Make sure your system is compatible with **CUDA 12.4**. If not, install [CuPy](https://docs.cupy.dev/en/stable/install.html) according to your current CUDA version.

## 🚀 Inference Example

```python
import os
from PIL import Image
import numpy as np
import matplotlib.pyplot as plt

from torchvision.transforms import Compose, ToTensor, Resize, Normalize
from utils.utils import denorm
from model.hub import MultiInputResShiftHub

model = MultiInputResShiftHub.from_pretrained("vfontech/Multiple-Input-Resshift-VFI").cuda()
model.eval()

img0_path = r"_data\example_images\frame1.png"
img2_path = r"_data\example_images\frame3.png"

mean = std = [0.5]*3
transforms = Compose([
    Resize((256, 448)),
    ToTensor(),
    Normalize(mean=mean, std=std),
])

img0 = transforms(Image.open(img0_path).convert("RGB")).unsqueeze(0).cuda()
img2 = transforms(Image.open(img2_path).convert("RGB")).unsqueeze(0).cuda()
tau = 0.5

img1 = model.reverse_process([img0, img2], tau)

plt.figure(figsize=(10, 5))
plt.subplot(1, 3, 1)
plt.imshow(denorm(img0, mean=mean, std=std).squeeze().permute(1, 2, 0).cpu().numpy())
plt.subplot(1, 3, 2)
plt.imshow(denorm(img1, mean=mean, std=std).squeeze().permute(1, 2, 0).cpu().numpy())
plt.subplot(1, 3, 3)
plt.imshow(denorm(img2, mean=mean, std=std).squeeze().permute(1, 2, 0).cpu().numpy())
plt.show()
```