The authors propose a terminal-fitted repair for classifier-free guidance (CFG) to address oversaturation and instability issues that occur at large guidance scales in diffusion and flow-matching samplers. By analyzing CFG through an asymptotic-preserving lens, they demonstrate that standard guidance causes the deterministic DDIM step to diverge as the minimum sigma approaches zero.

  • The repair replaces the standard CFG term w(r-1) with r^(1+w)-r on the guidance direction, requiring only one coefficient and zero extra network function evaluations.
  • This modification removes the sigma_min-divergent blow-up and achieves first-order accuracy against the exact guided flow as sigma_min goes to zero.
  • On learned CIFAR-10 checkpoints and Stable Diffusion 1.5 DDIM, the method acts as a high-guidance stabilizer that cuts residual amplification and saturation.
  • The approach yields point-FID wins over standard CFG on tested grids while preserving classifier-proxy target accuracy in hard-cell blocks.

The authors note that this repair is not a universal image-quality improvement but serves specifically to stabilize high-guidance sampling without additional computational cost.