# Any-to-Any Generation via Composable Diffusion

**Authors:** Zineng Tang, Ziyi Yang, Chenguang Zhu, Michael Zeng, Mohit Bansal  
**arXiv:** [2305.11846](https://arxiv.org/abs/2305.11846)  
**Date:** 2023-05-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.11846v1/

## Key points

- Composable Diffusion (CoDi) is a generative model that can generate any combination of output modalities from any combination of input modalities.
- CoDi can generate multiple modalities in parallel and is highly customizable and flexible.
- The model achieves strong joint-modality generation quality and outperforms or is on par with the unimodal state-of-the-art for single-modality synthesis.
- CoDi's ability to align modalities in both the input and output space allows it 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 that involves building a shared multimodal space by bridging alignment in the diffusion process, enabling synchronized generation of intertwined modalities such as temporally aligned video and audio.
- CoDi's ability to generate multiple modalities simultaneously makes it ideal for real-world applications where multiple modalities coexist and interact.
- CoDi consistently outperformed other state-of-the-art models across different tasks such as image captioning, text-to-image synthesis, and audio captioning.
- The project page with demonstrations and code is available at https://codi-gen.github.io/.

## Summary

Composable Diffusion (CoDi) is a groundbreaking generative model that can generate any combination of output modalities such as language, image, video or audio from any combination of input modalities. Unlike existing generative AI systems that are limited to a subset of modalities like text or image, CoDi can generate multiple modalities in parallel and is highly customizable and flexible. The model achieves strong joint-modality generation quality and outperforms or is on par with the unimodal state-of-the-art for single-modality synthesis. One of the key features of CoDi is its ability to align modalities in both the input and output space. This allows the model to freely condition on any input combination and generate any group of modalities even if they are not present in the training data. To achieve this alignment, CoDi employs a novel composable generation strategy that involves building a shared multimodal space by bridging alignment in the diffusion process. This enables the synchronized generation of intertwined modalities such as temporally aligned video and audio. While other models may be restricted in their real-world applicability where multiple modalities coexist and interact, CoDi's ability to generate multiple modalities simultaneously makes it ideal for real-world applications. Additionally, while one could chain together modality-specific generative models in a multi-step generation setting, this approach would be inherently limited by each step's generation power. A serial multi-step process can also be cumbersome and slow. CoDi's impressive performance has been demonstrated through various evaluations across different tasks such as image captioning, text-to-image synthesis and audio captioning where it consistently outperformed other state-of-the art models. Overall, Composable Diffusion represents an exciting breakthrough in generative modeling technology that has significant potential for real world applications across various industries. The project page with demonstrations and code is available at https://codi-gen.github.io/.