Conditional diffusion models can produce high‑quality images that closely follow a text prompt, yet samples drawn from different random seeds often look visually alike, limiting creative exploration. Prior techniques increase diversity but usually require extra optimization steps or sampling overhead, and strong perturbations may cause distortions or break prompt alignment. We introduce a training‑free approach that exploits the local pullback geometry induced by the mapping from the conditioning‑embedding space to the image space. By computing the pullback metric of the Jacobian and extracting its leading eigenvectors, we obtain directions of maximal output sensitivity. During sampling, we steer each random trajectory along these pullback directions, encouraging distinct generations while preserving consistency with the conditioning signal. Experiments on MS‑COCO and BrushNet demonstrate that the method markedly improves diversity for prompt‑conditioned image synthesis and inpainting, without sacrificing image fidelity or semantic alignment.
Review