BusterX

[🚀🤗BusterX++] [📦️🤗GenBuster-Bench++] [📦️🤗GenBuster-Bench] [📦️🤗GenBuster-Unified] [📦️🤗GenBuster-200K] [📦️🤗GenBuster-200K-mini] [🔥📜BusterX++ paper] [🔥📜BusterX paper]

---

Overview

BusterX is a family of MLLM-based methods for detecting and explaining AI-generated content. The project includes:

• BusterX++: Towards Unified Cross-Modal AI-Generated Content Detection and Explanation with MLLM
• BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation

Important

We release the 2026/06 Revised Edition of BusterX and BusterX++. Base models have been updated to the latest Qwen3.5, and the datasets and benchmarks have been updated.

Installation

We use uv for environment management. The default CUDA version is 12.x; check pyproject.toml if you need a different CUDA version.

# pip install uv
uv sync

Dataset & Benchmark Preparation

Log in to Hugging Face (make sure you have access to the gated datasets and models), then download and build all datasets:

make login
make fetch_data
make build_all_datasets

Available datasets and benchmarks:

Name	Description
GenBuster-200K / 200K-mini	Large-scale video training data
GenBuster-Unified	Image + video training data
GenBuster-Bench	Video-only evaluation benchmark
GenBuster-Bench++	Cross-modal (image + video) evaluation benchmark

Quick Start

Default settings target 8×H100 80GB. Adjust hyperparameters as needed.

Evaluate BusterX++ on GenBuster-Bench

bash scripts/eval_genbuster_bench.sh l8cv/BusterX-plusplus

For cross-modal evaluation on GenBuster-Bench++:

bash scripts/eval_genbuster_bench_plusplus.sh l8cv/BusterX-plusplus

Train with DAPO

We recommend server rollout mode for stability and efficiency. Start the rollout server first:

bash scripts/rollout.sh

Then launch training:

bash scripts/train_dapo.sh

GSPO training is also supported:

bash scripts/train_gspo.sh

Custom Benchmark Evaluation

make build_custom_benchmark
bash scripts/eval_custom_benchmark.sh <model>

Citation

@article{wen2025busterxunifiedcrossmodalaigenerated,
    title={BusterX++: Towards Unified Cross-Modal AI-Generated Content Detection and Explanation with MLLM},
    author={Haiquan Wen and Tianxiao Li and Zhenglin Huang and Yiwei He and Guangliang Cheng},
    journal={Arxiv},
    year={2025},
}

@article{wen2025busterxmllmpoweredaigeneratedvideo,
    title={BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation},
    author={Haiquan Wen and Yiwei He and Zhenglin Huang and Tianxiao Li and Zihan Yu and Xingru Huang and Lu Qi and Baoyuan Wu and Xiangtai Li and Guangliang Cheng},
    journal={Arxiv},
    year={2025},
}

Acknowledgement

This work is built upon ms-swift. Thanks for their excellent work!

License

This project is licensed under the BSD 3-Clause License — see LICENSE for details.