Deep dive into the STARFlow architecture, combining the expressiveness of autoregressive transformers with the efficiency of normalizing flows.
STARFlow integrates a Transformer backbone into a Normalizing Flow framework.
A Normalizing Flow is a sequence of invertible transformations that maps a simple distribution (like a Gaussian) to a complex data distribution (like images).
You can go from Data -> Noise (Training) and Noise -> Data (Generation) using the exact same math.
Unlike GANs, you can calculate the exact probability of a data point.
Standard flows struggle to model long-range dependencies (global structure in an image). STARFlow solves this by using a Transformer to parameterize the flow's transformations.
The model predicts the distribution of the next "patch" of the image based on previous patches.
Instead of just predicting a pixel value, it predicts the parameters of a flow transformation that generates the pixel.