How can we distill large text-to-image models without losing the diversity and controllability of the original teacher model? Excited to share our new ICML 2026 paper: Restoring Initial Noise Sensitivity in Text-to-Image Distillation via Geometric Alignment. #icml #generation Fast T2I distillation often maintains high image quality but makes the student much less sensitive to initial noise. This hurts an important property of generative models: changing the input noise should meaningfully change the output, which matters for diversity, layout variation, and noise-driven control. In this work, we trace this issue to standard distillation objectives that enforce pointwise output matching and unintentionally flatten the teacher’s local input-output geometry. We propose GAD (Geometry-Aware Distillation), a plug-and-play objective that preserves the teacher’s local functional behavior by aligning responses to noise perturbations through Jacobian-vector products. ✅ Better diversity ✅ Stronger layout/control transfer ✅ Preserved visual fidelity ✅ No extra inference cost Paper: arxiv.org/abs/2606.01651 Code: github.com/Hannah1102/GAD