justdubit/audiovisual_translation_dub

Audiovisual Translation Dubbing Dataset

A paired video dataset for training video dubbing and lip-sync models, as described in the paper "JUST-DUB-IT: Video Dubbing via Joint Audio-Visual Diffusion".

📄 Project: justdubit.github.io
💻 Code: github.com/justdubit/just-dub-it
🤗 Model: huggingface.co/justdubit/justdubit

Dataset Description

This dataset contains paired video samples for training synchronized audio-video dubbing models. Each sample includes:

• Target video: The dubbed output with translated speech and synchronized lip movements
• Reference video: The original source video used as conditioning
• Face segmentation masks: (Optional) Binary masks highlighting facial regions for masked loss training
• Text captions: Structured prompts with speaker description, language, and dialogue

Intended Use

• Training LoRA adapters for video dubbing tasks
• Fine-tuning audio-video models for lip-sync generation
• Research in multimodal translation and dubbing

Dataset Structure

The dataset is organized as follows:

• audiovisual_translation_dub/
  • dataset.json - Metadata file with captions and paths
  • target/ - Dubbed output videos
    • 001.mp4
    • 002.mp4
    • ...
  • reference/ - Original source videos
    • 001.mp4
    • 002.mp4
    • ...
  • face_masks/ - (Optional) Face segmentation masks
    • 001.mp4
    • 002.mp4
    • ...

Data Format

The dataset.json file contains entries in the following format:

[ { "caption": "The woman is speaking English, saying: 'Hello, how are you today?'", "media_path": "videos/target_001.mp4", "reference_path": "references/reference_001.mp4", "mask_path": "masks/mask_001.mp4" } ]