MicroMat-3k

# ZIM: Zero-Shot Image Matting for Anything

## Introduction

🚀 Introducing ZIM: Zero-Shot Image Matting – A Step Beyond SAM! 🚀

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. That’s where ZIM comes in.

🌟 What is ZIM? 🌟

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.

🔍 Get Started with ZIM 🔍

Ready to elevate your AI projects with unmatched matting quality? Access ZIM on our [project page](https://naver-ai.github.io/ZIM/), [Arxiv](https://huggingface.co/papers/2411.00626), and [Github](https://github.com/naver-ai/ZIM).

## Installation

```bash
pip install zim_anything
```

or

```bash
git clone https://github.com/naver-ai/ZIM.git
cd ZIM; pip install -e .
```

## Usage

1. Make the directory `zim_vit_l_2092`.
2. Download the [encoder](https://huggingface.co/naver-iv/zim-anything-vitl/resolve/main/zim_vit_l_2092/encoder.onnx?download=true) weight and [decoder](https://huggingface.co/naver-iv/zim-anything-vitl/resolve/main/zim_vit_l_2092/decoder.onnx?download=true) weight.
3. Put them under the `zim_vit_b_2092` directory.

```python
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:

```bibtex
@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}
}

# ZIM：万物零样本图像抠图（Zero-Shot Image Matting for Anything）

## 简介

🚀 推出ZIM：万物零样本图像抠图——超越SAM的全新方案！🚀

尽管SAM（分段任意模型，Segment Anything Model）凭借其在多领域的广泛应用重新定义了零样本分割，但在生成高精度、细粒度掩码方面往往存在不足。这正是ZIM的用武之地。

🌟 何为ZIM？🌟

ZIM（零样本图像抠图，Zero-Shot Image Matting）是一款突破性模型，旨在确立高精度抠图的全新标准，同时保留出色的零样本泛化能力。与SAM类似，ZIM能够在零样本范式下适配多样化的数据集与目标对象，但其表现更胜一筹，可生成捕捉复杂细节的高精度细粒度掩码。

🔍 快速上手ZIM 🔍

准备好以无与伦比的抠图质量升级你的AI项目了吗？可通过以下渠道获取ZIM：[项目主页](https://naver-ai.github.io/ZIM/)、[Arxiv论文](https://huggingface.co/papers/2411.00626)与[Github仓库](https://github.com/naver-ai/ZIM)。

## 安装

bash
pip install zim_anything

或

bash
git clone https://github.com/naver-ai/ZIM.git
cd ZIM; pip install -e .

## 使用方法

1. 创建`zim_vit_l_2092`目录。
2. 下载[编码器权重](https://huggingface.co/naver-iv/zim-anything-vitl/resolve/main/zim_vit_l_2092/encoder.onnx?download=true)与[解码器权重](https://huggingface.co/naver-iv/zim-anything-vitl/resolve/main/zim_vit_l_2092/decoder.onnx?download=true)。
3. 将权重文件放入`zim_vit_b_2092`目录中。

python
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>)

## 引用

若你认为本项目对你有帮助，请引用以下文献：

bibtex
@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}
}