zaqxsw0526/MMDF

X-AVDT: Audio-Visual Cross-Attention for Robust Deepfake Detection

Youngseo Kim · Kwan Yun · Seokhyeon Hong · Sihun Cha · Colette Suhjung Koo · Junyong Noh

Visual Media Lab @ KAIST
CVPR 2026

This repository contains the official implementation of the paper "X-AVDT: Audio-Visual Cross-Attention for Robust Deepfake Detection".

TL;DR : audio-visual cross-attention with diffusion inversion features, a more robust detector against unseen deepfake generators.

Installation

conda create -n x-avdt python=3.10 -y
conda activate x-avdt
pip install -r requirements.txt

ffmpeg is required for video preprocessing.

Download Pretrained Models

The Hallo feature extraction uses the original Hallo repository. Pretrained weights are placed under hallo/pretrained_models.

hallo/pretrained_models/

The easiest setup is to clone the official pretrained model bundle from Hugging Face:

cd hallo
git lfs install
git clone https://huggingface.co/fudan-generative-ai/hallo pretrained_models
cd ..

If downloading files manually, organize them as follows:

hallo/pretrained_models/
  audio_separator/
    download_checks.json
    mdx_model_data.json
    vr_model_data.json
    Kim_Vocal_2.onnx
  face_analysis/
    models/
      face_landmarker_v2_with_blendshapes.task
      1k3d68.onnx
      2d106det.onnx
      genderage.onnx
      glintr100.onnx
      scrfd_10g_bnkps.onnx
  hallo/
    net.pth
  motion_module/
    mm_sd_v15_v2.ckpt
  sd-vae-ft-mse/
    config.json
    diffusion_pytorch_model.safetensors
  stable-diffusion-v1-5/
    unet/
      config.json
      diffusion_pytorch_model.safetensors
  wav2vec/
    wav2vec2-base-960h/
      config.json
      feature_extractor_config.json
      model.safetensors
      preprocessor_config.json
      special_tokens_map.json
      tokenizer_config.json
      vocab.json

These paths match hallo/configs/inference/default.yaml.

Dataset

The MMDF can be downloaded from Hugging Face.

After feature extraction, training expects this layout for both real and fake roots:

<root>/<split>/<label>/<model_id>/<clip_id>/
  original/*.pt
  inverted/*.pt
  reconstructed/*.pt
  residual/*.pt
  attn_feat/*.pt

Feature Extraction

If your inputs are raw videos, first convert them into frame folders and wav files:

python hallo/preprocess_videos.py extract-frames \
  --video_dir /path/to/videos \
  --frames_dir /path/to/frames \
  --duration 5 \
  --fps 25 \
  --size 512 512

Then run Hallo feature extraction. This produces whole-clip outputs such as original.mp4, inverted.mp4, reconstructed.mp4, residual.mp4, attn_map.mp4, and attn_feat.pt for each clip:

python hallo/extract_features.py \
  --frames_dir /path/to/frames \
  --output_dir /path/to/hallo_features

Finally, pack the whole-clip feature outputs into the training layout. This step slices each clip into 16-frame .pt chunks:

python hallo/preprocess_videos.py pack-features \
  --feature_dir /path/to/hallo_features \
  --output_dir /path/to/pt \

Training

python train/train.py --data_dir /path/to/pt/ 

Evaluation

python train/evaluate.py --data_dir /path/to/pt/ --ckpt results/x_avdt/model_best.pt 

Citation

@article{kim2026x,
  title={X-AVDT: Audio-Visual Cross-Attention for Robust Deepfake Detection},
  author={Kim, Youngseo and Yun, Kwan and Hong, Seokhyeon and Cha, Sihun and Koo, Colette Suhjung and Noh, Junyong},
  journal={arXiv preprint arXiv:2603.08483},
  year={2026}
}