https://github.com/bytedance-seed/deepflow
[ICCV 2025] Deeply Supervised Flow-Based Generative Models
https://github.com/bytedance-seed/deepflow
Last synced: 3 months ago
JSON representation
[ICCV 2025] Deeply Supervised Flow-Based Generative Models
- Host: GitHub
- URL: https://github.com/bytedance-seed/deepflow
- Owner: ByteDance-Seed
- License: other
- Created: 2025-04-27T11:32:15.000Z (about 1 year ago)
- Default Branch: main
- Last Pushed: 2025-06-26T03:13:42.000Z (12 months ago)
- Last Synced: 2025-06-26T04:21:41.639Z (12 months ago)
- Language: Python
- Homepage: https://deepflow-project.github.io/
- Size: 722 KB
- Stars: 12
- Watchers: 3
- Forks: 2
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
👋 Hi, everyone!
We are ByteDance Seed team.
You can get to know us better through the following channels👇
# Deeply Supervised Flow-Based Generative Models
### [ICCV 2025]
Inkyu Shin ·
Chenglin Yang ·
Liang-Chieh Chen
DeepFlow is a novel flow-based generation framework that enhances velocity representation through inter layer communication using deep supervision and acceleration mechanism.
DeepFlow converges 8 times faster on ImageNet with equivalent performance and further reduces FID by 2.6 while halving training time compared to previous flow based models without a classifier free guidance.
# News
[2025/06/25]🎉DeepFlow is accepted to ICCV 2025.
[2025/05/07]🔥We release training code, sampling code, and checkpoints of DeepFlow.
# Introduction
DeepFlow incorporates deep supervision by evenly adding velocity prediction within transformer blocks, further enhanced by the
proposed Velocity Alignment block (VeRA).
# Getting started
## Installation
```shell
pip install -r requirements.txt
```
## Dataset
We conducted our experiments using [ImageNet](https://www.kaggle.com/competitions/imagenet-object-localization-challenge/data).
Kindly refer to preprocessing guide from [REPA](https://github.com/sihyun-yu/REPA/tree/main/preprocessing) to preprocess the dataset.
## Training
For training DeepFlow-XL/2-3T, use below training script.
```shell
bash script/xlarge/train.sh
```
where you can set following options:
- `--tg-upper-bound`: time-gap between adjacent branches.
- `--weighting`: time-step sampling during training.
- `--df-idxs`: key transformer layers where deep supervision is applied. (currently, equally-splitted).
- `--ssl-align`: whether to use SSL align (need to set up `--enc-type` as well).
- `--legacy-scaling`: whether to use soft-cap for scaling factor in "velocity_modulation" (enable this when ssl-align is enabled for reproducibility).
## Generation
For generative samples from DeepFlow-XL/2-3T, use below evaluation script.
```shell
bash script/xlarge/eval.sh
```
where you need to set specific path for checkpoint that will be evaluated.
Please follow [ADM evaluation](https://github.com/openai/guided-diffusion/tree/main/evaluations) for obtaining FID score.
## Model Zoo
We release checkpoint of DeepFlow trained on ImageNet for your reference.
| Dataset | Model | SSL align | Training Epochs | Link | FID (wo/ CFG) | FID (w/ CFG) |
| ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- |
| ImageNet-256px | DeepFlow-XL/2-3T | X | 400 | [checkpoint](https://drive.google.com/file/d/1QXTDkeVRgfvBvyLkJcu-YJ_2pDxM73Ec/view?usp=sharing) | 7.2 | 1.97 |
| ImageNet-256px | DeepFlow-XL/2-3T | O | 400 | [checkpoint](https://drive.google.com/file/d/1ALMz0Cdm1XSZYWTPSVEVR-sfYXsKk6Zl/view?usp=sharing) | 5.0 | 1.77 |
# License
The code and model weights are licensed under CC-BY-NC. See LICENSE.txt for details.
# Citation
If you find DeepFlow useful for your research and applications, feel free to give us a star ⭐ or cite us using:
```bibtex
@article{shin2025deeply,
title={Deeply Supervised Flow-Based Generative Models},
author={Shin, Inkyu and Yang, Chenglin and Chen, Liang-Chieh},
journal={arXiv preprint arXiv:2503.14494},
year={2025}
}
```
## Acknowledgement
[DiT](https://github.com/facebookresearch/DiT)
[SiT](https://github.com/willisma/SiT)
[REPA](https://github.com/sihyun-yu/REPA)
# About [ByteDance Seed Team](https://team.doubao.com/)
Founded in 2023, ByteDance Seed Team is dedicated to crafting the industry's most advanced AI foundation models. The team aspires to become a world-class research team and make significant contributions to the advancement of science and society.