Infinity introduces a novel approach to visual autoregressive modeling, enabling the generation of high-resolution, photorealistic images. By employing a bitwise token prediction framework, it achieves faster inference and improved generation quality compared to existing models.
Key Features
Bitwise Token Prediction: Utilizes a bitwise multi-scale residual quantizer and infinite-vocabulary classifier to enhance token prediction efficiency.
Bitwise Self-Correction: Implements a mechanism to mitigate train-test discrepancies, improving image generation consistency.
High-Resolution Generation: Capable of generating 1024×1024 images with significantly reduced inference time.
Pretrained Models: Provides access to various model checkpoints, including Infinity-2B and Infinity-8B, for diverse applications.
Interactive Demos: Offers Jupyter Notebooks for hands-on experimentation and understanding of model behavior.
Android DevHub
 (1).jpg)
.png)
.png)
.png)
.png)
.png)
