---
license: apache-2.0
pretty_name: WEAVE
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*
task_categories:
- any-to-any
- image-to-image
- text-to-image
- visual-question-answering
language:
- en
tags:
  - image
  - image-editing
  - multimodal
library_name: datasets
---

# 🧵 WEAVE

<div align="center">

[![arXiv](https://img.shields.io/badge/arXiv-2511.11434-b31b1b.svg)](https://arxiv.org/abs/2511.11434)
[![Dataset](https://img.shields.io/badge/🤗%20Huggingface-Dataset-yellow)](https://huggingface.co/datasets/WeiChow/Weave/)
[![Checkpoint](https://img.shields.io/badge/🤗%20Huggingface-CKPT-blue)](https://huggingface.co/WeiChow/Bagel-weave)
[![GitHub](https://img.shields.io/badge/GitHub-Repo-181717?logo=github)](https://github.com/weichow23/weave)
[![Page](https://img.shields.io/badge/🏠%20Home-Page-9cf.svg)](https://weichow23.github.io/weave/)

</div>

> 📌 This is the official repository for the dataset presented in:  
> **Weave: A Benchmark for Evaluating Multimodal Editing Models**

## 📊 Dataset Structure

WEAVE consists of two main components:
- 🔹 **WEAVE-100k** (Training Set): Stored in `data/` folder in parquet format
- 🔹 **WEAVEBench** (Test Set): Stored in `test/` folder in both zip and json formats

<img src="asset/teaser.png" alt="Test Set Categories" width="700" />

## 🚀 WEAVE-100k

<img src="asset/pipe.png" alt="Test Set Categories" width="700" />

WEAVE-100k is generated through four sophisticated pipelines and multiple validation stages using state-of-the-art VLMs and image generation models:

- ✨ Leverages cutting-edge models including GPT-4.1, Nano Banana, and SeeDream 4.0
- ✨ Used to fine-tune [Bagel](https://huggingface.co/WeiChow/Bagel-weave), which achieved superior results on multiple benchmarks including GenEval and GeditBench
- ✨ For more details, please refer to our paper: [![arXiv](https://img.shields.io/badge/arXiv-2511.15738-b31b1b.svg)](https://arxiv.org/abs/2511.15738)

## 🧪 WEAVEBench

WEAVEBench is manually designed and curated, featuring 16 diverse categories of editing tasks.

<img src="asset/test.png" alt="Test Set Categories" width="700" />

### 📝 Test Set File Format (`test.json`)

> ⚠️ **Note**: `Image #1` references the first image, starting from 1. This represents the image index, not the conversation turn. It corresponds to the first image in the `images` array (`images[0]`).
>
> When using multi-turn conversations, each number index should be replaced once with `Image #{idx}<image>\n`. For single-turn, simply replace directly.

```json
{
    "domain": "string",
    "images": [],
    "chats": []
}
```

## 🔍 Want to Test on WEAVEBench?

Please refer to our code repository: [![GitHub](https://img.shields.io/badge/GitHub-Repo-181717?logo=github)](https://github.com/weichow23/weave)

## ✍️ Citation

```bibtex
@article{weave2024,
  title={Weave: A Benchmark for Evaluating Multimodal Editing Models},
  author={Wei Chow et al.},
  journal={arXiv preprint arXiv:2511.15738},
  year={2024}
}
```