---
title: "Tel Aviv University Cuts Image Generation Steps in Half by Rethinking the Prior"
subtitle: "Researchers sidestep the intractable OT problem entirely by redesigning the prior, cutting trajectory curvature in half and boosting few-step image generation quality."
url: "https://alphasignal.ai/news/tel-aviv-university-cuts-image-generation-steps-in-half-by-rethinking-the-prior"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-06-02T00:00:00.000Z"
updated: "2026-06-15T00:54:12.732Z"
type: "Paper"
topics:
  - "Image"
  - "Llms"
subtopics:
  - "Image Generation"
upvotes: 68
source_url: "https://arxiv.org/abs/2606.04092"
comments: 0
truncated: true
---

# Tel Aviv University Cuts Image Generation Steps in Half by Rethinking the Prior

> Researchers sidestep the intractable OT problem entirely by redesigning the prior, cutting trajectory curvature in half and boosting few-step image generation quality.

Flow matching has quietly become the backbone of modern image generators. [Stable Diffusion 3, FLUX, and other leading models](https://arxiv.org/abs/2606.04092) all rely on it. The core idea is elegant: instead of the noisy, multi-step denoising of diffusion models, flow matching trains a continuous-time model by directly matching a velocity field, enabling efficient deterministic sampling with far fewer steps while preserving quality. But there is a persistent bottleneck that has frustrated the field for years, and a new paper from Tel Aviv University and Reichman University proposes a surprisingly clean fix.

## The Curvature Problem

The holy grail in flow matching is getting the model to learn perfectly straight trajectories from noise to data. Straight paths mean you can take fewer integration steps (each step requires a full neural network forward pass, called a Neural Function Evaluation, or NFE), which directly translates to faster generation. When prior-data pairs are coupled via optimal transport (OT), the learned trajectories are straight and non-crossing, enabling fast, even single-step, generation.

The problem is that computing the true OT coupling is computationally intractable at the scale of real images. OT coupling means finding the mathematically optimal way to match every noise sample to every data sample so that the total "transport cost" is minimized. Existing methods attempt to solve the OT problem, at the cost of persistent bias or significant overhead. Minibatch OT approximates the coupling over small batches and introduces bias; semi-discrete OT methods like AlignFlow require expensive precomputed assignments that don't scale beyond small datasets. The field has been stuck trying to solve a hard problem, when perhaps the problem itself could be reframed.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [Optimal Transport Flow Matching by Design (arXiv)](https://arxiv.org/abs/2606.04092)
- [Designing Optimal Transport Flows (arXiv HTML)](https://arxiv.org/html/2606.04092)
- [Flow Matching in Generative Modeling (Emergent Mind)](https://www.emergentmind.com/topics/flow-matching-fm)
- [COT-FM: Cluster-wise Optimal Transport Flow Matching](https://embodiedai-ntu.github.io/cotfm/)
- [Flow-Multi: A Flow-Matching Multi-Reward Framework](https://pmc.ncbi.nlm.nih.gov/articles/PMC12943997/)