Search papers, labs, and topics across Lattice.
This paper introduces Signed Rectified Flow (Signed RF), a novel generative modeling framework that generalizes Rectified Flow to effectively manage signed measures by promoting certain distributions while suppressing others. The method is significant because it allows for the incorporation of negative information and exclusion constraints, addressing challenges in generative modeling that involve undesirable outputs. Key results demonstrate that Signed RF enhances the fidelity-diversity trade-off on ImageNet, mitigates nearest-neighbor similarity in anti-memorization tasks, and reduces unwanted nudity in image generation while maintaining aesthetic quality.
Signed Rectified Flow not only improves generative fidelity but also effectively suppresses unwanted outputs, reshaping the landscape of controlled generation.
We introduce Signed Rectified Flow (Signed RF), a generalization of Rectified Flow that targets the signed measure $\pi^{sign} = (1+\alpha)\pi^+ - \alpha\pi^-$, where $\alpha>0$, $\pi^+$ is the distribution to promote, and $\pi^-$ is the distribution to suppress. Although direct sampling from a signed measure is not well-defined, Signed RF induces a valid generative process that concentrates probability in regions where the signed measure is positive while provably excluding regions dominated by its negative component. It therefore provides a principled framework for incorporating negative information and exclusion constraints into generative modeling. We analyze the signed continuity equation underlying Signed RF and use a charged-particle interpretation to explain how negative mass forms exclusion barriers. This theory further motivates practical adaptive guidance algorithms. Across several applications, Signed RF improves the fidelity-diversity trade-off on ImageNet, reduces nearest-neighbor similarity in anti-memorization experiments, and reduces nudity induced by adversarial prompts in Stable Diffusion 3.5 while preserving CLIP and aesthetic scores.
