EPFL-VILAB/A2A-Video-models

A2A-Video: Modeling the World Across Time and Modalities

Muhammad Uzair Khattak, Won Jun Kim, Reza Abbassi, Michael Murphy, Albias Havolli, Amir Zadeh, Chuan Li, Oğuzhan Fatih Kar, Muhammad Ferjad Naeem, Andrei Atanov, Roman Bachmann, Federico Tombari, Amir Zamir

EPFL, Google, Lambda

Official pre-trained checkpoints for A2A-Video, an any-to-any multimodal model for the video domain that allows flexible traversal across both modality and time axes.

🌐 Project page | 📄 Paper | 💻 Code | 🤗 Tokenizers | 🤗 Demo | BibTeX

This repo holds the main A2A-Video model checkpoints. We release two versions:

Model	Params	File	Config
A2A-Video-L	705M	main_model/A2A-Video-L.pth	Config
A2A-Video-Pred-L	705M	main_model/A2A-Video-Pred-L.pth	Config

A2A-Video-Pred is an A2A-Video model whose pretraining input-output mixture is biased toward predicting video from sparse modalities (e.g. caption, transcription) and toward forecasting tasks -- it can still perform all other any-to-any tasks.

For the tokenizers these checkpoints depend on, see EPFL-VILAB/A2A-Video-tokenizers.

Usage

The easiest way to get started is the 🤗 Space demo, or the notebooks in the code repo, which download these checkpoints automatically:

from notebooks import pipeline
from notebooks.display_utils import video_grid

tokenizers = pipeline.load_tokenizers()                            # downloads the 7 tokenizers from the Hub on first call
model = pipeline.load_model(tokenizers, model_name="A2A-Video_Pred_L")  # downloads the model from the Hub on first call

example = pipeline.load_example(pipeline.list_examples()[0])

results = pipeline.generate(
    example, input_modalities=["rgb"],
    chain=pipeline.CHAIN_PRESETS["rgb_to_others"]["chain"],
    tokenizers=tokenizers, model=model,
)
video_grid(results, cols=5, width=170, title="rgb -> everything else")

See the main repo README for installation, training, and evaluation instructions.

License

These model weights are released under the Sample Code license as found in LICENSE_WEIGHTS.

Citation

If you find this repository useful, please consider citing:

@article{a2avideo2026,
    title={{A2A-Video: Modeling the World Across Time and Modalities}},
    author={Khattak, Muhammad Uzair and Kim, Won Jun and Abbassi, Reza and Havolli, Albias and Murphy, Michael and Zadeh, Amir and Li, Chuan and Kar, O\u{g}uzhan Fatih and Naeem, Muhammad Ferjad and Atanov, Andrei and Bachmann, Roman and Tombari, Federico and Zamir, Amir},
    journal={arXiv preprint},
    year={2026},
}