DriftXpress paper claims 6.68x faster drifting-model training

One-step image generators like drifting models solve one problem, iterative sampling, by pushing much of the compute into training. A new arxiv preprint from Ali Falahati, Elliot Creager, Gautam Kamath, and Shubhankar Mohapatra tries to buy some of that training cost back with an approximation trick, and reports numbers big enough to be worth writing down.

The setup: standard drifting models repeatedly evaluate an attraction kernel against every training point. DriftXpress replaces exact kernel interactions using a Nyström feature map built from landmark points, so the field is approximated in a low-rank space, while generated-sample repulsion is kept exact for stability. The authors say this preserves what they call the attraction-repulsion structure of the original drifting field.

The numbers are the interesting part. On SVHN, throughput reportedly climbs from 2,307 to 15,402 images per second, a 6.68x speedup, with FID nudging from 2.94 to 3.11. On CIFAR10 the paper observes a 6.63x throughput speedup and a small FID improvement, from 5.64 to 5.52. Higher-resolution or more-diverse datasets fare less well: 2.95x on CIFAR100 and 2.64x on ImageNet, where FID moves in the wrong direction, from 8.83 to 9.21. The authors report that the optimal landmark ratio sits between 0.0256 and 0.1024, or 128 to 512 landmarks per class.

The honest caveats are what you would expect from a single preprint. The speedup is a training-throughput number rather than a full wall-clock or dollar cost, and the paper does not benchmark against distilled diffusion or consistency models, which are the more mainstream ways to get one-step samplers today. The ImageNet result also suggests the quality gap widens as you scale, which matters if anyone wants to try this at text-to-image resolutions rather than small research benchmarks.

If it holds up, the practical read is that drifting models get cheaper to iterate on, which is the kind of thing that decides whether a niche paradigm ever leaves the niche. A PyTorch implementation is available on GitHub for anyone who wants to reproduce.