README
FlowBind: Efficient Any-to-Any Generation with Bidirectional Flows
Overview
FlowBind is an efficient flow-based multimodal generation model with bidirectional flows.
Try it out in our demo! : demo.ipynb
Model Description
The following files are required for inference and training:
| Filename | Description |
|---|---|
model.pth | Main FlowBind model weights. Includes three drift networks, one for each modality (Text, Image, Audio). |
config.json | Configuration file containing model hyperparameters and architecture details. |
gemmae_3.pt | Weights for the Text Decoder (Gemma). |
clap.pt | Weights for the Audio Encoder (CLAP). |
Data Preparation
We utilize the following datasets for training and evaluation.
- Text-Image:
- LAION-COCO (Filtered by aesthetic scores)
- Flickr30k
- Text-Audio:
- Audio-Image:
Note: Due to copyright and licensing restrictions, we cannot provide the raw training datasets directly. Please download the data from the official links provided above.
Citation
If you find our work helpful, please cite:
@inproceedings{
cha2026flowbind,
title={FlowBind: Efficient Any-to-Any Generation with Bidirectional Flows},
author={Yeonwoo Cha and Semin Kim and Jinhyeon Kwon and Seunghoon Hong},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=7DeARTwvwL}
}
Contact
For any inquiries, please contact Yeonwoo Cha at ckdusdn03@kaist.ac.kr.
Contributors
13 commits
Similar repos
README
FlowBind: Efficient Any-to-Any Generation with Bidirectional Flows
Overview
FlowBind is an efficient flow-based multimodal generation model with bidirectional flows.
Try it out in our demo! : demo.ipynb
Model Description
The following files are required for inference and training:
| Filename | Description |
|---|---|
model.pth | Main FlowBind model weights. Includes three drift networks, one for each modality (Text, Image, Audio). |
config.json | Configuration file containing model hyperparameters and architecture details. |
gemmae_3.pt | Weights for the Text Decoder (Gemma). |
clap.pt | Weights for the Audio Encoder (CLAP). |
Data Preparation
We utilize the following datasets for training and evaluation.
- Text-Image:
- LAION-COCO (Filtered by aesthetic scores)
- Flickr30k
- Text-Audio:
- Audio-Image:
Note: Due to copyright and licensing restrictions, we cannot provide the raw training datasets directly. Please download the data from the official links provided above.
Citation
If you find our work helpful, please cite:
@inproceedings{
cha2026flowbind,
title={FlowBind: Efficient Any-to-Any Generation with Bidirectional Flows},
author={Yeonwoo Cha and Semin Kim and Jinhyeon Kwon and Seunghoon Hong},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=7DeARTwvwL}
}
Contact
For any inquiries, please contact Yeonwoo Cha at ckdusdn03@kaist.ac.kr.
Similar repos
Linked in READMEs
Contributors
13 commits