Sihyeon Kim1 Seunghun Lee2 Vikas Singh3,§ Hyunwoo J. Kim2,§
1Korea University 2KAIST 3University of Wisconsin-Madison
Few-step ODE solvers for diffusion and flow models, whether analytic (iPNDM, DPM-Solver++) or learned (DyWeight, S4S), build each step as a scalar-weighted combination of buffered velocities, so the update can never leave the linear span of the velocity buffer. SpanLift keeps a fixed base solver as an in-span prior and adds a lightweight spatial residual operator over the state and velocity buffer that corrects the update beyond this span. It is trained by endpoint teacher matching, keeps the pretrained backbone frozen, and adds no extra model NFEs. With only 3 NFE, SpanLift improves FID from 8.16 to 5.69 on CIFAR-10 and from 17.37 to 11.83 on ImageNet 64×64.
🚧 Coming soon.
@inproceedings{kim2026spanlift,
title = {Learning to Solve Generative {ODEs} Beyond the Linear Span},
author = {Kim, Sihyeon and Lee, Seunghun and Singh, Vikas and Kim, Hyunwoo J.},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026}
}