4M: Massively Multimodal Masked Modeling

David Mizrahi*, Roman Bachmann*, Oğuzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, Amir Zamir

Official implementation and pre-trained models for "4M: Massively Multimodal Masked Modeling" (NeurIPS 2023).

Website | Paper | GitHub

4M is a framework for training "any-to-any" foundation models, using tokenization and masking to scale to many diverse modalities. Models trained using 4M can perform a wide range of vision tasks, transfer well to unseen tasks and modalities, and are flexible and steerable multimodal generative models.

Installation

For install instructions, please see https://github.com/apple/ml-4m.

Usage

The depth tokenizer can be loaded from Hugging Face Hub as follows:

from fourm.vq.vqvae import DiVAE
tok_rgb = DiVAE.from_pretrained('EPFL-VILAB/4M_tokenizers_depth_8k_224-448')

Please see https://github.com/apple/ml-4m/blob/main/README_TOKENIZATION.md for more detailed instructions and https://github.com/apple/ml-4m for other tokenizer and 4M model checkpoints.

Safetensors checkpoints are hosted under https://huggingface.co/EPFL-VILAB/4M.

Citation

If you find this repository helpful, please consider citing our work:

@inproceedings{mizrahi20234m,
    title={{4M}: Massively Multimodal Masked Modeling},
    author={David Mizrahi and Roman Bachmann and O{\u{g}}uzhan Fatih Kar and Teresa Yeo and Mingfei Gao and Afshin Dehghan and Amir Zamir},
    booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
    year={2023},
}

License

The model weights in this repository are released under the Sample Code license as found in the LICENSE file.