README.md

---
license: apache-2.0
language:
- en
tags:
- video-editing
- text-to-video
- diffusion-transformer
- sparse-attention
- wan
- tilelang
- triton
pipeline_tag: image-to-video
library_name: pytorch
base_model:
- Wan-AI/Wan2.2-T2V-A14B
---

<div align="center">

# LIVEditor-14B

### Lightning Unified Video Editing via In-Context Sparse Attention

**Shitong Shao** · **Zikai Zhou** · **Haopeng Li** · **Yingwei Song** · **Wenliang Zhong** · **Lichen Bai** · **Zeke Xie**

[![Project Page](https://img.shields.io/badge/Project-Page-green)](https://xie-lab-ml.github.io/liveditor-page/)
[![Paper](https://img.shields.io/badge/Paper-arXiv-red)](https://arxiv.org/abs/2605.04569)
[![Code](https://img.shields.io/badge/Code-GitHub-blue)](https://github.com/xie-lab-ml/Lightning-Unified-Video-Editor-via-In-Context-Sparse-Attention)
[![Model](https://img.shields.io/badge/Model-HuggingFace-yellow)](https://huggingface.co/sst12345/liveditor)

</div>

<p align="center">
  <img src="./assets/live_visualization.jpg" alt="LIVEditor-14B teaser" width="92%">
</p>

## Overview

**LIVEditor-14B** is a unified video editing model built for fast in-context video editing. It introduces **In-Context Sparse Attention (ISA)**, a lightweight sparse attention mechanism that retrieves relevant source-video context blocks instead of applying dense full attention over all source and generated video tokens.

The model is designed to preserve the editing quality of in-context full-attention video editing while substantially reducing attention latency.

## Highlights

- **Unified video editing**: one editor for diverse text-guided video editing scenarios.
- **In-Context Sparse Attention**: retrieves only the most relevant source-video blocks for each query block.
- **Training-free acceleration block**: ISA can be plugged into the diffusion transformer attention backend.
- **Efficient sparse kernels**: supports both **TileLang** and **Triton** implementations.
- **Strong speedup**: the paper reports up to **2.8× faster** attention than FlashAttention-2 at 65K tokens on RTX 4090.

## Demo

<table>
<tr>
<th>Source Video</th>
<th>LIVEditor-14B Output (TileLang)</th>
<th>LIVEditor-14B Output (Triton)</th>
</tr>
<tr>
<td><img src="https://huggingface.co/sst12345/liveditor/resolve/main/assets/input.gif" alt="Source video preview" width="260"></td>
<td><img src="https://huggingface.co/sst12345/liveditor/resolve/main/assets/output_tilelang.gif" alt="TileLang output preview" width="260"></td>
<td><img src="https://huggingface.co/sst12345/liveditor/resolve/main/assets/output_triton.gif" alt="Triton output preview" width="260"></td>
</tr>
</table>

MP4 downloads: [source](https://huggingface.co/sst12345/liveditor/resolve/main/assets/input.mp4) · [TileLang output](https://huggingface.co/sst12345/liveditor/resolve/main/assets/output_tilelang.mp4) · [Triton output](https://huggingface.co/sst12345/liveditor/resolve/main/assets/output_triton.mp4)

More qualitative comparisons are available on the [project page](https://xie-lab-ml.github.io/liveditor-page/).

## Method

<p align="center">
  <img src="./assets/in_context_sparse_attention.png" alt="In-Context Sparse Attention" width="92%">
</p>

LIVEditor-14B stores compressed key/value representations of the source video, computes block-wise relevance scores, retrieves top-*k* source blocks, and applies sparse piecewise attention for efficient in-context editing. Query blocks with sharper attention patterns can use full FlashAttention, while diffuse blocks use the sparse Top-K path.

## Quick Start

Clone the code repository:

```bash
git clone https://github.com/xie-lab-ml/Lightning-Unified-Video-Editor-via-In-Context-Sparse-Attention.git
cd Lightning-Unified-Video-Editor-via-In-Context-Sparse-Attention
pip install -r requirements.txt
```

Download the LIVEditor-14B checkpoint:

```bash
pip install huggingface_hub
huggingface-cli download sst12345/liveditor liveditor_ckpt.bin --local-dir .
```

Run inference:

```bash
python inference.py \
  --config inference.yaml \
  --checkpoint liveditor_ckpt.bin \
  --input assets/input.mp4 \
  --prompt "Add a small golden crown with delicate jewels on top of the girl's head..." \
  --output result.mp4
```

## Model Files

| File | Description |
|---|---|
| `liveditor_ckpt.bin` | LIVEditor-14B fine-tuned checkpoint |
| `assets/live_visualization.jpg` | Teaser image for the model card |
| `assets/in_context_sparse_attention.png` | Method overview |
| `assets/input.mp4` | Example input video |
| `assets/output_tilelang.mp4` | Example output using TileLang backend |
| `assets/output_triton.mp4` | Example output using Triton backend |
| `assets/input.gif` | Browser-friendly source preview |
| `assets/output_tilelang.gif` | Browser-friendly TileLang preview |
| `assets/output_triton.gif` | Browser-friendly Triton preview |

## Citation

```bibtex
@article{shao2026liveditor,
  title={LIVEditor-14B: Lightning Unified Video Editing via In-Context Sparse Attention},
  author={Shao, Shitong and Zhou, Zikai and Li, Haopeng and Song, Yingwei and Zhong, Wenliang and Bai, Lichen and Xie, Zeke},
  journal={arXiv preprint arXiv:2605.04569},
  year={2026}
}
```