jeasco / sefm

Separating Anything from Image in Context

Separating Anything from Image in Context (in submission)

💡 Highlight

• 😍 😍 SeFM is a novel generalist foundation model for separating anything from images.
• 😍 😍 Only a handy demonstration example is required.
• 😍 😍 Exhibiting groundbreaking generalization capability to open-world unseen visual components.

🏷️ TODO

• Release inference code and demo.
• Release checkpoints.
• Release training codes.

🎖️ Framework Architecture

Overall Framework of SeFM

🛠️ Installation

The model is built in PyTorch 1.8.0 and tested on Ubuntu 16.04 environment (Python3.7, CUDA9.0, cuDNN7.5).

1. Clone the SeFM repository from GitHub:

git clone https://github.com/Jeasco/SeFM
cd SeFM

2. Install the required dependencies and SeFM:

pip3 install -r requirements.txt
python3 -m pip install -e .

Training

bash sefm_training.sh

Evaluation

1. Download the pre-trained model and place it in ./checkpoints/

2. Place the test image in ./test/

3. Run

python in_context_inference_demo.py

4. Visual results will be saved in results