New AI Image Trick Makes Generators Follow Your Prompts Better

Conditioning diffusion models usually means retraining or complex modifications. This research offers a plug-in alternative: train one unconditional joint score network, then steer generation at inference with a simple correction term. That term reveals exactly how the condition shapes the target distribution. The authors derive explicit conditional SDEs and probability-flow ODEs, making different conditional samplers directly comparable. They also add a log-Fokker–Planck residual regularization to shrink the gap between ODE and SDE sampling, and experiments show it improves deterministic sampling while delivering competitive conditional image generation.