One-step flow samplers lose their uncertainty: Isthmus rewires only the time sampler and brings brain-MRI calibration back to multi-step levels
Synopsis
When one-step flow maps serve as posterior samplers, the spread of their samples is the uncertainty they are meant to carry, yet the standard one-step sampler loses that spread and decorrelates from the 64-step samples of the same network. Isthmus spreads training time pairs along the whole edge of the time triangle that joins the one-step pair to the diagonal, so the one-step map's error is bounded by the self-consistency residual. On brain MRI, one-step calibration returns to 0.96 and correlation with multi-step samples to 0.93, and the agreement holds when training is extended to three times its length.
Interpretation
The sampler of a one-step flow map loses the uncertainty the model itself carries and drifts apart from the same network's multi-step samples. The 64-step sampler of the same network is close to calibrated, while the standard one-step samples are already under-dispersed, with calibration 0.67 and correlation 0.58 against the 64-step samples. On IXI brain MRI slices, the same weights are sampled in one step and in 64 steps; the 64-step sampler is near calibrated across several methods, while one-step sampling departs from it.
The cause lies in where training places its start-end time pairs on the time triangle, not in the loss itself. Only pairs on the diagonal receive targets from data, and the pair used for one-step sampling sits on the landing edge, where the standard sampler places no training pairs, so the objective neither determines the one-step map uniquely nor anchors its training. Moving the same share of training pairs to the corner point, the launch strip, the corner square or the landing strip, only the option covering the landing edge is both consistent with 64 steps and calibrated, reaching 0.96 on IXI.
Isthmus is a mixture time sampler that changes only the time sampler, leaving the network, the loss and the inference code untouched. By default it replaces a training example's time pair with a jump from a uniform start that lands on the data, and these new pairs are trained with the unchanged mean-flow target. Across brain and knee MRI, CelebA face inpainting, and unconditional generation on CIFAR-10 and ImageNet, the same repair improves one-step sampling under several flow-map objectives at unchanged per-iteration cost.
Perspective
The demonstrated scope is posterior sampling and generation: IXI brain MRI slices, fastMRI knee data, CelebA face inpainting, and unconditional generation on CIFAR-10 and ImageNet, all trained with a conditional U-Net and the mean-flow loss. The intended readers are researchers and engineers using few-step flow models for sampling or inverse problems.
Bridging bounds the error of the one-step map but does not by itself calibrate the samples, and discrete targets such as step doubling in shortcut models fall outside this analysis. For objectives that differentiate in the end time, the bridge moves to the launch edge, and placing pairs on the wrong edge lets one-step variance grow to about 2.68 times the 64-step variance.
