Bert Based Text Conditioning With Classifier Free Guidance

via “classifier-free guidance with dynamic thresholding for text alignment control”

Implementation of Imagen, Google's Text-to-Image Neural Network, in Pytorch

Unique: Combines classifier-free guidance with dynamic thresholding (percentile-based clipping) rather than fixed-value thresholding, enabling automatic adaptation to different prompt difficulties and model scales without per-prompt manual tuning

vs others: Provides better artifact prevention than fixed-threshold guidance and requires no separate classifier network unlike traditional guidance methods, reducing training complexity while improving robustness across diverse prompts