Developed by a team at ByteDance (yes, TikTok, your favorite social network), this model called PuLID will allow you to create custom images from existing photos, all in the blink of an eye.
Based on the principle of contrastive alignment, PuLID offers ultra-fast and high-quality identity customization. It does this by using an architecture that learns key characteristics of a source identity (photos of you) on the fly and efficiently translates them to new target images (AI-generated images). The result is unique visuals generated in just a few seconds, while preserving the consistency of the original photos.
Well, ok, it may sound a bit complex, but it’s super simple to use. If you are lazy, you can go directly to Huggingface or, for the bravest, install it on your machine. All you need is good old PyTorch (version 2.0 minimum, please) and a few lines of code to start the training.
PuLID (Pure and Lightning ID Customization via Contrastive Alignment) works by using machine learning techniques to align latent representations by comparing pairs of images or identities. The goal is to maximize similarity for similar identities and minimize similarity for different identities. By adjusting these representations with contrastive alignment, PuLID makes it possible to create unique images with great precision and speed.
If you’re in comm and already have some cool concept art of a character but would like to see what it would look like in different environments or with different graphic styles, don’t worry! You throw your images into PuLID with the right instructions and you’re done. You’ll get a whole bunch of stylish variations of your character while keeping their face recognizable.
The ByteDance team has thought of everything: PuLID is 100% open-source and available on GitHub. You can therefore tinker with the code as you see fit for your needs. There are even tutorials and examples to help you get to grips with the book quickly.
And for the most impatient among you, here is a little installation tutorial to start playing with PuLID:
Prerequisites:
- Python >= 3.7
- PyTorch >= 2.0 (download here)
- Anaconda (download here) or Miniconda (download here)
Installation steps:
- Clone the PuLID repository:
git clone https://github.com/ToTheBeginning/PuLID.git
cd PuLID- Create and activate the conda environment:
conda create --name pulid python=3.10
conda activate pulid- Install the dependencies:
pip install -r requirements.txt- Install PyTorch: Follow the instructions on the PyTorch website to install the compatible version for your system. For example, for CUDA 11.7:
conda install pytorch torchvision torchaudio cudatoolkit=11.7 -c pytorch- Launch the application:
python app.pyTo learn more about PuLID and retrieve the source code, visit the GitHub repo.
Come on, I’ll let you have fun with your new toy. Many thanks to Lorenper for the info. Thanks to you, we’ll be able to customize our avatars like never before.
