What is Classifier-Free Guidance?

Classifier-Free Guidance or CFG is a technique used in text-to-image and text-to-video diffusion models to make the generated outputs more closely follow the given prompt.

Normally, diffusion models can generate images by conditioning on text (using the prompt) or unconditionally (without any prompt).

Classifier-free guidance works by combining these two cases: during training, the model sometimes ignores the prompt so it learns both conditional and unconditional generation.

At inference time, we run the model twice-once with the prompt and once without-and then adjust the result by pushing the conditional output further away from the unconditional one.

This effectively "amplifies" the influence of the prompt, guiding the model to align more closely to the described content.

By tuning the guidance scale (alpha), you can control and balance between creativity and alignment to the prompt.

Want to Learn More About ML/Al? Refine your skills with our Weekly Al Newsletter-educational, easy to understand, mathematically explained, and completely free (link in bio O).

C: Welch Labs

Join our Al community for more posts like this ‪@Giffah_Alexander‬

#machinelearning #deeplearning #mathematics #math #statistics #datascience #computerscience #diffusion #generativeai