Awesome

This repo is the official implementation of "Bilateral Reference for High-Resolution Dichotomous Image Segmentation" (CAAI AIR 2024).

[!note] We need more GPU resources to push forward the performance of BiRefNet, especially on video tasks and more efficient model designs on higher-resolution images. If you are happy to cooperate, please contact me at zhengpeng0108@gmail.com.

News :newspaper:

• Dec 5, 2024: Fix the bug of using torch.compile in latest PyTorch versions (2.5.1) and the slow iteration in FP16 training with accelerate (set as default).
• Nov 28, 2024: Congrats to students @Nankai University employed BiRefNet to build their project and won the provincial gold medal and national bronze medal on the China International College Students’ Innovation Competition 2024.
• Oct 26, 2024: We added the guideline of conducting fine-tuning on custom data with existing weights.
• Oct 6, 2024: We uploaded the BiRefNet-matting model for general trimap-free matting use.
• Sep 24, 2024: We uploaded the BiRefNet_lite-2K model, which takes inputs in a much higher resolution (2560x1440). We also added the notebook for inference on videos.
• Sep 7, 2024: Thanks to Freepik for supporting me with GPUs for more extensive experiments, especially on BiRefNet for 2K inference!
• Aug 30, 2024: We uploaded notebooks in tutorials to run the inference and ONNX conversion locally.
• Aug 23, 2024: Our BiRefNet is now officially released online on CAAI AIR journal. And thanks to the press release.
• Aug 19, 2024: We uploaded the ONNX model files of all weights in the GitHub release and GDrive folder. Check out the ONNX conversion part in model zoo for more details.
• Jul 30, 2024: Thanks to @not-lain for his kind efforts in adding BiRefNet to the official huggingface.js repo.
• Jul 28, 2024: We released the Colab demo for box-guided segmentation.
• Jul 15, 2024: We deployed our BiRefNet on Hugging Face Models for users to easily load it in one line code.
• Jun 21, 2024: We released and uploaded the Chinese version of our original paper to my GDrive.
• May 28, 2024: We hold a model zoo with well-trained weights of our BiRefNet in different sizes and for different tasks, including general use, matting segmentation, DIS, HRSOD, COD, etc.
• May 7, 2024: We also released the Colab demo for multiple images inference. Many thanks to @rishabh063 for his support on it.
• Apr 9, 2024: Thanks to Features and Labels Inc. for deploying a cool online BiRefNet inference API and providing me with strong GPU resources for 4 months on more extensive experiments!
• Mar 7, 2024: We released BiRefNet codes, the well-trained weights for all tasks in the original papers, and all related stuff in my GDrive folder. Meanwhile, we also deployed our BiRefNet on Hugging Face Spaces for easier online use and released the Colab demo for inference and evaluation.
• Jan 7, 2024: We released our paper on arXiv.

:rocket: Load BiRefNet in ONE LINE by HuggingFace, check more:

from transformers import AutoModelForImageSegmentation
birefnet = AutoModelForImageSegmentation.from_pretrained('zhengpeng7/BiRefNet', trust_remote_code=True)

:flight_arrival: Inference Partner:

We are really happy to collaborate with FAL to deploy the inference API of BiRefNet. You can access this service via the link below:

• https://fal.ai/models/fal-ai/birefnet

Our BiRefNet has achieved SOTA on many similar HR tasks:

DIS:

COD:

HRSOD:

Try our online demos for inference:

• Inference and evaluation of your given weights:
• Online Inference with GUI with adjustable resolutions:
• Online Multiple Images Inference on Colab:

Model Zoo

For more general use of our BiRefNet, I extended the original academic one to more general ones for better real-life application.

Datasets and datasets are suggested to be downloaded from official pages. But you can also download the packaged ones: DIS, HRSOD, COD, Backbones.

Find performances (almost all metrics) of all models in the exp-TASK_SETTINGS folders in [stuff].

• Given box guidance:

Screenshot from the original paper. All tests are conducted on a single A100 GPU.

<img src="https://drive.google.com/thumbnail?id=1mTfSD_qt-rFO1t8DRQcyIa5cgWLf1w2-&sz=h300" /> <img src="https://drive.google.com/thumbnail?id=1F_OURIWILVe4u1rSz-aqt6ur__bAef25&sz=h300" />

We converted from .pth weights files to .onnx files.
We referred a lot to the Kazuhito00/BiRefNet-ONNX-Sample, many thanks to @Kazuhito00.

• Check our Colab demo for ONNX conversion or the notebook file for local running, where you can do the conversion/inference by yourself and find all relevant info.
• As tested, BiRefNets with SwinL (default backbone) cost ~90% more time (the inference costs ~165ms on an A100 GPU) using ONNX files. Meanwhile, BiRefNets with SwinT (lightweight) cost ~75% more time (the inference costs ~93.8ms on an A100 GPU) using ONNX files. Input resolution is 1024x1024 as default.
• The results of the original pth files and the converted onnx files are slightly different, which is acceptable.
• Pay attention to the compatibility among onnxruntime-gpu, CUDA, and CUDNN (we use torch==2.0.1, cuda=11.8 here).

Third-Party Creations

We found there've been some 3rd party applications based on our BiRefNet. Many thanks for their contribution to the community!
Choose the one you like to try with clicks instead of codes:

1. Applications:

   • Thanks tin2tin/2D_Asset_Generator: this project combined BiRefNet and FLUX as a Blender add-on for "AI generating 2D cutout assets for ex. previz".

     https://github.com/user-attachments/assets/6cce7ca7-7817-4406-b6c4-6d4e8c414ed4

   • Thanks camenduru/text-behind-tost: this project employed BiRefNet to extract foreground subjects and add texts between the subjects and background, which looks amazing especially for videos. Check their tweets for more examples.

     <p align="center"><img src="https://github.com/user-attachments/assets/9969dd10-38a8-4cf2-a6c7-5b11f074b9b4" height="300"/></p>

   • Thanks briaai/RMBG-2.0: this project trained BiRefNet with their high-quality private data, which brings improvement on the DIS task. Note that their weights are for only non-commercial use and are not aware of transparency due to training in the DIS task setting, which focuses only on predicting binary masks.

     <p align="center"><img src="https://huggingface.co/briaai/RMBG-2.0/resolve/main/t4.png" height="300"/></p>

   • Thanks lldacing/ComfyUI_BiRefNet_ll: this project further upgrade the ComfyUI node for BiRefNet with both our latest weights and the legacy ones.

     <p align="center"><img src="https://github.com/lldacing/ComfyUI_BiRefNet_ll/raw/main/doc/video.gif" height="300"/></p>

   • Thanks MoonHugo/ComfyUI-BiRefNet-Hugo: this project further upgrade the ComfyUI node for BiRefNet with our latest weights.

     <p align="center"><img src="https://github.com/MoonHugo/ComfyUI-BiRefNet-Hugo/raw/main/assets/demo4.gif" height="300"/></p>

   • Thanks lbq779660843/BiRefNet-Tensorrt and yuanyang1991/birefnet_tensorrt: they both provided the project to convert BiRefNet to TensorRT, which is faster and better for deployment. Their repos offer solid local establishment (Win and Linux) and colab demo, respectively. And @yuanyang1991 kindly offered the comparison among the inference efficiency of naive PyTorch, ONNX, and TensorRT on an RTX 4080S:

• Thanks dimitribarbot/sd-webui-birefnet: this project allows to add a BiRefNet section to the original Stable Diffusion WebUI's Extras tab.

  <p align="center"><img src="https://drive.google.com/thumbnail?id=159bLXI71FWh4ZsHTvc-wApSN9ytVRmua&sz=w1620" /></p>

• Thanks fal.ai/birefnet: this project on fal.ai encapsulates BiRefNet online with more useful options in UI and API to call the model.

  <p align="center"><img src="https://drive.google.com/thumbnail?id=1rNk81YV_Pzb2GykrzfGvX6T7KBXR0wrA&sz=w1620" /></p>

• Thanks ZHO-ZHO-ZHO/ComfyUI-BiRefNet-ZHO: this project further improves the UI for BiRefNet in ComfyUI, especially for video data.

  <p align="center"><img src="https://drive.google.com/thumbnail?id=1GOqEreyS7ENzTPN0RqxEjaA76RpMlkYM&sz=w1620" /></p>

  https://github.com/ZhengPeng7/BiRefNet/assets/25921713/3a1c7ab2-9847-4dac-8935-43a2d3cd2671

• Thanks viperyl/ComfyUI-BiRefNet: this project packs BiRefNet as ComfyUI nodes, and makes this SOTA model easier use for everyone.

  <p align="center"><img src="https://drive.google.com/thumbnail?id=1KfxCQUUa2y9T-aysEaeVVjCUt3Z0zSkL&sz=w1620" /></p>

• Thanks Rishabh for offering a demo for the easier multiple images inference on colab.

2. More Visual Comparisons

   • Thanks twitter.com/ZHOZHO672070 for the comparison with more background-removal methods in images:

     <img src="https://drive.google.com/thumbnail?id=1nvVIFt_Ezs-crPSQxUDqkUBz598fTe63&sz=w1620" />

   • Thanks twitter.com/toyxyz3 for the comparison with more background-removal methods in videos:

   https://github.com/ZhengPeng7/BiRefNet/assets/25921713/40136198-01cc-4106-81f9-81c985f02e31

   https://github.com/ZhengPeng7/BiRefNet/assets/25921713/1a32860c-0893-49dd-b557-c2e35a83c160

Usage

Environment Setup

# PyTorch==2.0.1+CUDA11.8 (or 2.5.1+CUDA12.4) is used for faster training (~40%) with compilation.
conda create -n birefnet python=3.11 -y && conda activate birefnet
pip install -r requirements.txt

Dataset Preparation

Download combined training / test sets I have organized well from: DIS--COD--HRSOD or the single official ones in the single_ones folder, or their official pages. You can also find the same ones on my BaiduDisk: DIS--COD--HRSOD.

Weights Preparation

Download backbone weights from my google-drive folder or their official pages.

Run

# Train & Test & Evaluation
./train_test.sh RUN_NAME GPU_NUMBERS_FOR_TRAINING GPU_NUMBERS_FOR_TEST
# Example: ./train_test.sh tmp-proj 0,1,2,3,4,5,6,7 0

# See train.sh / test.sh for only training / test-evaluation.
# After the evaluation, run `gen_best_ep.py` to select the best ckpt from a specific metric (you choose it from Sm, wFm, HCE (DIS only)).

:pen: Fine-tuning on Custom Data

Suppose you have some custom data, fine-tuning on it tends to bring improvement.

1. Pre-requisites: you have put your datasets in the path ${data_root_dir}/TASK_NAME/DATASET_NAME. For example, ${data_root_dir}/DIS5K/DIS-TR and ${data_root_dir}/General/TR-HRSOD, where im and gt are both in each dataset folder.
2. Change an existing task to your custom one: replace all 'General' (with single quotes) in the whole project with your custom task name as the screenshot of vscode given below shows:<img src="https://drive.google.com/thumbnail?id=1J6gzTmrVnQsmtt3hi6ch3ZrH7Op9PKSB&sz=w400" />
3. Adapt settings:
   • sys_home_dir: path to the root folder, which contains codes / datasets / weights / ... -- project folder / data folder / backbone weights folder are ${sys_home_dir}/codes/dis/BiRefNet / ${sys_home_dir}/datasets/dis/General / ${sys_home_dir}/weights/cv/swin_xxx, respectively.
   • testsets: your validation set.
   • training_set: your training set.
   • lambdas_pix_last: adapt the weights of different losses if you want, especially for the difference between segmentation (classification task) and matting (regression task).
4. Use existing weights: if you want to use some existing weights to fine-tune that model, please refer to the resume argument in train.py. Attention: the epoch of training continues from the epochs the weights file name indicates (e.g., 244 in BiRefNet-general-epoch_244.pth), instead of 1. So, if you want to fine-tune 50 more epochs, please specify the epochs as 294. \#Epochs, \#last epochs for validation, and validation step are set in train.sh.
5. Good luck to your training :) If you still have questions, feel free to leave issues (recommended way) or contact me.

Well-trained weights:

Download the BiRefNet-{TASK}-{EPOCH}.pth from [stuff] and the release page of this repo. Info of the corresponding (predicted_maps/performance/training_log) weights can be also found in folders like exp-BiRefNet-{TASK_SETTINGS} in the same directory.

You can also download the weights from the release of this repo.

The results might be a bit different from those in the original paper, you can see them in the eval_results-BiRefNet-{TASK_SETTINGS} folder in each exp-xx, we will update them in the following days. Due to the very high cost I used (A100-80G x 8), which many people cannot afford (including myself....), I re-trained BiRefNet on a single A100-40G only and achieved the performance on the same level (even better). It means you can directly train the model on a single GPU with 36.5G+ memory. BTW, 5.5G GPU memory is needed for inference in 1024x1024. (I personally paid a lot for renting an A100-40G to re-train BiRefNet on the three tasks... T_T. Hope it can help you.)

But if you have more and more powerful GPUs, you can set GPU IDs and increase the batch size in config.py to accelerate the training. We have made all these kinds of things adaptive in scripts to seamlessly switch between single-card training and multi-card training. Enjoy it :)

Some of my messages:

This project was originally built for DIS only. But after the updates one by one, I made it larger and larger with many functions embedded together. Finally, you can use it for any binary image segmentation tasks, such as DIS/COD/SOD, medical image segmentation, anomaly segmentation, etc. You can eaily open/close below things (usually in config.py):

• Multi-GPU training: open/close with one variable.
• Backbone choices: Swin_v1, PVT_v2, ConvNets, ...
• Weighted losses: BCE, IoU, SSIM, MAE, Reg, ...
• Training tricks: multi-scale supervision, freezing backbone, multi-scale input...
• Data collator: loading all in memory, smooth combination of different datasets for combined training and test.
• ... I really hope you enjoy this project and use it in more works to achieve new SOTAs.

Quantitative Results

Qualitative Results

Acknowledgement:

Many of my thanks to the companies / institutes below.

• FAL.
• Freepik.
• Redmond.ai.
• Alibaba-ICBU.

Citation

@article{zheng2024birefnet,
  title={Bilateral Reference for High-Resolution Dichotomous Image Segmentation},
  author={Zheng, Peng and Gao, Dehong and Fan, Deng-Ping and Liu, Li and Laaksonen, Jorma and Ouyang, Wanli and Sebe, Nicu},
  journal={CAAI Artificial Intelligence Research},
  volume = {3},
  pages = {9150038},
  year={2024}
}

Contact

Any questions, discussions, or even complaints, feel free to leave issues here (recommended) or send me e-mails (zhengpeng0108@gmail.com) or book a meeting with me: calendly.com/zhengpeng0108/30min. You can also join the Discord Group (https://discord.gg/d9NN5sgFrq) or QQ Group (https://qm.qq.com/q/y6WPy7WOIK) if you want to talk a lot publicly.