Datasets:
image
imagewidth (px) 1.14k
2.04k
|
---|
ZIM: Zero-Shot Image Matting for Anything
Introduction
๐ Introducing ZIM: Zero-Shot Image Matting โ A Step Beyond SAM! ๐
While SAM (Segment Anything Model) has redefined zero-shot segmentation with broad applications across multiple fields, it often falls short in delivering high-precision, fine-grained masks. Thatโs where ZIM comes in.
๐ What is ZIM? ๐
ZIM (Zero-Shot Image Matting) is a groundbreaking model developed to set a new standard in precision matting while maintaining strong zero-shot capabilities. Like SAM, ZIM can generalize across diverse datasets and objects in a zero-shot paradigm. But ZIM goes beyond, delivering highly accurate, fine-grained masks that capture intricate details.
๐ Get Started with ZIM ๐
Ready to elevate your AI projects with unmatched matting quality? Access ZIM on our project page, Arxiv, and Github.
Installation
pip install zim_anything
or
git clone https://github.com/naver-ai/ZIM.git
cd ZIM; pip install -e .
Usage
- Make the directory
zim_vit_l_2092
. - Download the encoder weight and decoder weight.
- Put them under the
zim_vit_b_2092
directory.
from zim_anything import zim_model_registry, ZimPredictor
backbone = "vit_l"
ckpt_p = "zim_vit_l_2092"
model = zim_model_registry[backbone](checkpoint=ckpt_p)
if torch.cuda.is_available():
model.cuda()
predictor = ZimPredictor(model)
predictor.set_image(<image>)
masks, _, _ = predictor.predict(<input_prompts>)
Citation
If you find this project useful, please consider citing:
@article{kim2024zim,
title={ZIM: Zero-Shot Image Matting for Anything},
author={Kim, Beomyoung and Shin, Chanyong and Jeong, Joonhyun and Jung, Hyungsik and Lee, Se-Yun and Chun, Sewhan and Hwang, Dong-Hyun and Yu, Joonsang},
journal={arXiv preprint arXiv:2411.00626},
year={2024}
}
- Downloads last month
- 542