ICML 2026 · arXiv:2608.12276

XYZFlow: Scaling Multidimensional Shortcut Flows for Efficient Generative Modeling

Jinxiu Liu · CUHKXuanming Liu · WestlakeKangfu Mei · Johns HopkinsYandong Wen · WestlakeWeiyang Liu · CUHK

A multidimensional flow-matching framework for high-fidelity, few-step image generation.

Explore the method ↓Read the paper ↗GitHub ↗

XYZFlow case study: quality and latency comparison

AbstractHigh-fidelity image generation faces a trade-off between speed and quality. Diffusion models produce strong visuals but require costly iterative sampling. Existing efficient methods mainly distill pretrained models into few-step samplers, a challenging process that depends heavily on teacher-model quality. In this paper, we introduce XYZFlow, a framework that rethinks efficient generation through multidimensional scaling of flow matching. Unlike single-step mappings, XYZFlow enhances expressivity by making probability paths more identifiable and learnable through structured multidimensional conditioning. We view autoregressive modeling as implicit flow straightening, where richer context reduces trajectory ambiguity. XYZFlow realizes this idea through two orthogonal dimensions: temporal scaling, which uses non-Markovian conditioning on the full denoising history; and spatial scaling, enabled by Next Shortcut Prediction, which sequentially generates patches using preceding patches’ denoising trajectories as priors. Experiments show that XYZFlow achieves state-of-the-art performance, with 7.2–8.5× teacher speedups and competitive FID, while Next Shortcut Prediction delivers superior quality-latency trade-offs over model scaling or step reduction.

The thesis

Efficiency is a
conditioning problem.

Instead of only shrinking models or sampling steps, XYZFlow increases the dimensionality of the context around each prediction—making the probability path straighter and more unique.

3–4inference steps in the low-step regime

2Dtemporal + spatial flow scaling

256²ImageNet evaluation resolution

SOTAcompetitive FID / quality–latency trade-off

Section 3 · The Proposed XYZFlow Framework

Method, as described
in the paper.

XYZFlow enhances the expressivity of flow models through multidimensional conditioning. The paper develops this idea from autoregressive flow enhancement to temporal and spatial conditioning.

3.1

Autoregressive Modeling as Flow Enhancement

The growing autoregressive context acts as progressively stronger constraints. It can reduce the variance of the conditional distribution and straighten the probability-flow path from noise to data, making the mapping more unique and deterministic.

3.2

A Motivating Observation from Progressive Constraint Strengthening

As more patches are generated, later patches become easier to sample: predictions become more precise, latent patch variance decreases, and denoising paths become straighter under stronger contextual conditioning.

3.3

Multidimensional Conditioning

The framework enhances flow matching along temporal and spatial dimensions. Temporal conditioning uses a patch’s complete generation history; spatial conditioning uses Next Shortcut Prediction to transfer the preceding patch’s denoising trajectory as guidance.

Paper figures

See the constraints
stack up.

Figure 1: ambiguous paths and Next Shortcut Prediction

Figure 1. Conventional one-shot denoising has overlapping, ambiguous probability paths. Next Shortcut Prediction instead transfers the preceding patch’s complete denoising trajectory as a prior for the next patch.

Figure 2. As the autoregressive context grows, constraints strengthen: later patches become easier to predict, their latent variance decreases, and their denoising paths become straighter.

Figure 3. XYZFlow combines intra-patch trajectory conditioning with cross-patch trajectory guidance to scale flow matching along temporal and spatial dimensions.

The takeaway

“Richer context reduces trajectory ambiguity”—so the next shortcut can be faster without becoming fragile.