Serena Yeung-Levy#1461
Generative models usually turn noise → data. But a lot of science needs unpaired data → data: untreated cells → post-intervention cells, low-redshift galaxies → high.
Flow matching can do this in principle — but its quality degrades sharply in high dimensions. The fix? Add more noise.
Stochastic Perturbations Improve Distribution-to-Distribution Generative Models 📍 #CVPR2026
Users share the paper on stochastic perturbations fixing flow matching degradation in high dimensions because it improves distribution-to-distribution generative models while thanking collaborators.
We propose to densify the supervision with stochastic injections - pre-train noise→target, then fine-tune onto source→target - jitter source samples with Gaussian noise - perturb the interpolant path
All three are training-time only, basically computationally free, no architecture changes, and you can sample with a plain ODE (no expensive SDE solver).
Beats vanilla flow by 13 FID & baselines by 9, across diverse scientific domains: 🔬🌄🛰️🩻🌌.
Paper: https://openaccess.thecvf.com/content/CVPR2026F/papers/Su_Stochastic_Perturbations_Improve_Distribution-to-Distribution_Generative_Models_CVPRF_2026_paper.pdf
Many thanks to my collaborators @Zhang_Yu_hui @linqi_zhou Rajesh Ranganath @yeung_levy!
We're presenting at #CVPR2026 Findings in Denver — come say hi at Poster #67, ExHall A, 7:30 – 9:00 AM tomorrow 👋
Why the problem? With finite samples on both ends, the training signal lives only along thin lines between source–target pairs.
In high dimensions they have vanishing cover over the space, so the velocity field suffers from sparse supervision and hence poor generalisation.