Classifier-Free Guidance
A powerful sampling technique used in diffusion models to dramatically increase the alignment between the generated image and the specific text condition.
Think of It Like This
Like an artist painting a portrait who constantly steps back to compare their current progress against the client's strict written description.
During training, the text prompt is randomly dropped out, forcing the model to learn both unconditional and conditional generation simultaneously. At inference time, the model computes both predictions and extrapolates away from the unconditional output toward the conditional one. This effectively forces the image to pay much closer attention to the prompt at the cost of some visual diversity.