Open-MAGVIT2
Open-MAGVIT2 offers advanced image reconstruction, with models ranging from 300M to 1.5B parameters, achieving 1.17 rFID on ImageNet 256×256.
English
English
What is Open-MAGVIT2?
Open-MAGVIT2 is an open-source series of autoregressive image generation models developed by Tencent's ARC Lab. The project includes models ranging from 300M to 1.5B parameters. It reproduces Google’s MAGVIT-v2 tokenizer and achieves advanced reconstruction performance on the ImageNet 256x256 dataset with a 1.17 rFID score.
Key Features:
Offers models from 300M to 1.5B parameters.
Replicates Google’s MAGVIT-v2 tokenizer.
Achieves 1.17 rFID on ImageNet 256x256.
Uses asymmetric tokenization to optimize large vocabulary prediction.
Introduces 'next sub-token prediction' to enhance image quality.
Supports training and testing on various hardware platforms.
Provides comprehensive documentation for easy setup and use.
Target Audience:
The project targets researchers, developers, and students interested in deep learning and image processing. It is ideal for professionals working on image reconstruction, style transfer, and image generation.
Use Cases:
High-quality image reconstruction to improve compression and transmission efficiency.
Style transfer tasks converting low-resolution images to high-resolution artistic styles.
Image synthesis for generating specific scenes or objects.
Getting Started:
1. Visit the GitHub page and clone or download the source code.
2. Install dependencies using pip based on the requirements.txt file.
3. Set up Python and CUDA environment as per the documentation.
4. Use provided training scripts and model configurations to start training.
5. Utilize trained models for image generation tasks, adjusting parameters to optimize results.
6. Fine-tune and optimize models for specific applications as needed.
