Why it matters

CFG is default conditioning. Understanding shapes SD + video diffusion.

Advertisement

The architecture

Predict noise conditioned + unconditioned.

Sample: (1+w) * cond - w * uncond.

CFG mathConditional predeps(x, c)Unconditional predeps(x)Guided(1+w) cond - w uncondw typically 5-15; higher = closer to condition but degraded quality
CFG.
Advertisement

How it works end to end

Ho + Salimans 2022.

Train with condition dropout.

Guidance scale w tunes fidelity vs diversity.

w = 7.5 typical for SD.