Title: Reinstating Initial Noise Sensitivity in Text-to-Image Distillation Through Geometric Alignment

Abstract: Generative distillation serves as a powerful mechanism for accelerating text-to-image (T2I) synthesis. By compressing multi-step generation trajectories into streamlined, few-step student models, it achieves significant speedups without compromising perceptual quality. Despite these advances, current methodologies tend to prioritize computational efficiency and output accuracy, frequently overlooking essential characteristics inherent to the original generation trajectory. This study highlights a critical oversight: the degradation of sensitivity to initial noise. This deficiency hampers downstream control techniques that depend on noise-based optimization and manipulation.

We attribute this problem to conventional distillation objectives, which focus on pointwise output alignment. Such approaches inadvertently flatten the input-output landscape, thereby erasing the teacher model’s local geometric structure. To remedy this, we introduce Geometry-Aware Distillation (GAD), a framework designed to preserve sensitivity by aligning the local functional behaviors of teacher and student models. GAD achieves this by matching Jacobian-vector products relative to input noise, allowing the student model to accurately replicate the teacher’s differential responses to perturbations. Our extensive evaluations across various T2I paradigms and noise-driven control tasks confirm that GAD effectively restores noise sensitivity and enhances diversity, all while sustaining high visual fidelity. The source code can be accessed at https://github.com/Hannah1102/GAD.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC