CoDi: Any-to-Any Generation via Composable Diffusion

Zineng Tang^1*, Ziyi Yang^2†, Chenguang Zhu², Michael Zeng², Mohit Bansal^1†

¹University of North Carolina at Chapel Hill, ²Microsoft Azure Cognitive Services Research ^* Work done at Microsoft internship and UNC. ^†Corresponding Authors

Introduction

We present Composable Diffusion (CoDi), a novel generative model capable of generating any combination of output modalities, such as language, image, video, or audio, from any combination of input modalities. Unlike existing generative AI systems, CoDi can generate multiple modalities in parallel and its input is not limited to a subset of modalities like text or image. Despite the absence of training datasets for many combinations of modalities, we propose to align modalities in both the input and output space. This allows CoDi to freely condition on any input combination and generate any group of modalities, even if they are not present in the training data. CoDi employs a novel composable generation strategy which involves building a shared multimodal space by bridging alignment in the diffusion process, enabling the synchronized generation of intertwined modalities, such as temporally aligned video and audio. Highly customizable and flexible, CoDi achieves strong joint-modality generation quality, and outperforms or is on par with the unimodal state-of-the-art for single-modality synthesis.

Installation

conda create -n CoDi python=3.8 # prepare an environment

pip install torch==1.12.1+cu116 torchaudio==0.12.1+cu116 torchvision==0.13.1+cu116 \
-f https://download.pytorch.org/whl/torch_stable.html # change cuda version higher pytorch version is doable

pip install -r requirement.txt

Inference

Download checkpoints from Hugginface Model

Model Parts	Huggingface Weights Address	fp16 weights
CoDi Encoders and VAEs	CoDi_encoders.pth	CoDi_encoders.pth
CoDi Text Diffuser	CoDi_text_diffuser.pth	CoDi_text_diffuser.pth
CoDi Audio Diffuser	CoDi_audio_diffuser_m.pth	CoDi_audio_diffuser_m.pth
CoDi Vision Diffuser	CoDi_video_diffuser_8frames.pth	CoDi_video_diffuser_8frames.pth

wget https://huggingface.co/ZinengTang/CoDi/resolve/main/CoDi_encoders.pth -P checkpoints/
wget https://huggingface.co/ZinengTang/CoDi/resolve/main/CoDi_text_diffuser.pth -P checkpoints/
wget https://huggingface.co/ZinengTang/CoDi/resolve/main/CoDi_audio_diffuser_m.pth -P checkpoints/
wget https://huggingface.co/ZinengTang/CoDi/resolve/main/CoDi_video_diffuser_8frames.pth -P checkpoints/

# Or fp16 weights
wget https://huggingface.co/ZinengTang/CoDi/resolve/main/checkpoints_fp16/CoDi_encoders.pth -P checkpoints/
wget https://huggingface.co/ZinengTang/CoDi/resolve/main/checkpoints_fp16/CoDi_text_diffuser.pth -P checkpoints/
wget https://huggingface.co/ZinengTang/CoDi/resolve/main/checkpoints_fp16/CoDi_audio_diffuser_m.pth -P checkpoints/
wget https://huggingface.co/ZinengTang/CoDi/resolve/main/checkpoints_fp16/CoDi_video_diffuser_8frames.pth -P checkpoints/

Run demo.ipynb

Citation

If you find our work useful, please consider citing:

@article{tang2023any,
  title={Any-to-Any Generation via Composable Diffusion},
  author={Tang, Zineng and Yang, Ziyi and Zhu, Chenguang and Zeng, Michael and Bansal, Mohit},
  journal={arXiv preprint arXiv:2305.11846},
  year={2023}
}

Reference

The code structure is based on Versatile Diffusion. The audio diffusion model is based on AudioLDM. The video diffusion model is partially based on Make-A-Video.

Usage and License Notices:

CoDi released in this repository is free to use for most circumstances under MIT license.

Contact

Zineng Tang (zn.tang.terran@gmail.com)