Spectral Guidance Enables Fast, Training-Free Control of Diffusion Models
Spectral Guidance leverages the intrinsic geometry of diffusion processes to project arbitrary conditioning signals—labels, CLIP embeddings, masks—onto the sampling trajectory without retraining. The method boosts conditional accuracy on CIFAR‑10 by 37 points over the strongest training‑free baseline while cutting sampling time by four‑fold, and reveals an optimal guidance window.