Straight Path Generative Modeling

This direction is merged into another direction: trajectory straightening, curved sampling trajectories. Its last version stays readable and citable here.

Version 1 of 1218 papers240 conceptsfrozen 2026-10-04

Standard diffusion models are notoriously slow because they treat the journey from noise to data like a winding mountain road. To reach the destination without flying off a cliff, a numerical solver must take hundreds of tiny, cautious steps. Recent research has shifted toward a more efficient geometry: turning that winding road into a straight highway. By "straightening" the probability flow, we can generate high-quality samples in as little as a single step.

The Cost of Curvature

In continuous-time generative models, we define a vector field that pushes a simple distribution (like Gaussian noise) toward a complex one (like the manifold of human faces). Traditionally, this vector field is learned through a process that pairs noise and data points somewhat arbitrarily.

As Xingchao Liu, Chengyue Gong, and Qiang Liu 0001 explain in Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow, these arbitrary pairings force the model to learn highly curved trajectories. If point A in the noise space is paired with point B in the data space, but the path between them crosses or tangles with other paths, the resulting "average" velocity field becomes non-linear.

Curvature is the enemy of speed. A curved path requires many Function Evaluations (NFE) because a linear approximation (like an Euler step) quickly accumulates error. To maintain fidelity, solvers must reduce their step size, leading to the high computational "tax" seen in modern diffusion models.

Straightening via Optimal Transport

The most effective way to reduce curvature is to ensure that the noise-to-data pairings are as "orderly" as possible. This is where Optimal Transport (OT) comes in. Instead of random pairing, researchers like Michal Klein and marco cuturi suggest using large-scale Sinkhorn couplings to align distributions.

In On Fitting Flow Models with Large Sinkhorn Couplings, the authors demonstrate that training on massive OT couplings—involving millions of points—results in significantly straighter trajectories. Similarly, Alireza Mousavi Hosseini and colleagues introduced Semidiscrete Flow Matching (SD-FM), which uses a global transport map to eliminate the "path crossing" that typically occurs when matching small random batches. By ensuring that the noise distribution is perfectly aligned with the data points, the model learns a velocity field that is nearly constant along each path.

Reaching the One-Step Limit

The ultimate goal of path straightening is "one-step generation," where a single Euler step is sufficient to produce a high-fidelity image. InstaFlow, developed by Xingchao Liu and Jianzhu Ma, achieved this by iteratively "rectifying" the flow of Stable Diffusion. By training a new model to follow the (now straighter) paths of a teacher model, they distilled the complex multi-step process into a 0.09-second inference step.

However, even with straight paths, neural network approximations can introduce subtle non-linearities. To combat this, Bo Chen and Chengyue Gong proposed HOMO (High-Order Matching for One-Step Shortcut Diffusion). Rather than just matching velocity, HOMO matches higher-order dynamics like acceleration and jerk. This allows the model to capture the underlying geometry of the data manifold with much higher precision, stabilizing the training of one-step "shortcut" models.

Adapting the Source to the Target

Another perspective on straightening involves changing the starting point. Standard models often try to force "square" data into "round" Gaussian noise holes. Jannis Chemseddine and Gabriele Steidl argue in Adapting Noise to Data that using data-adaptive noise distributions can significantly shorten transport paths.

This concept is extended to conditional generation in CAR-Flow by Chen Chen and colleagues. By shifting the source noise distribution based on the target class (e.g., moving the noise closer to the "dog" region of the data manifold when generating a dog), the model reduces the distance and complexity the flow must travel.

Beyond Images: Robotics and Science

The benefits of straight paths extend far beyond image synthesis. In robotics, where latency is critical, ActionFlow and COT Policy use Flow Matching and Optimal Transport to enable real-time control. By ensuring the mathematical paths from noise to robot actions are straight, these models can operate at 100Hz with 10x less computation than standard diffusion policies.

In the natural sciences, Rectified Flows for Fast Multiscale Fluid Flow Modeling by Victor Armegioiu and Siddhartha Mishra applies these same principles to chaotic fluid simulations, achieving 22x speedups. Whether the target is a pixel, a protein backbone (Simon V Mathis et al., Flows, straight but not so fast), or a therapeutic peptide (Hao Qian et al., POTFlow), the intuition remains the same: the shortest distance between noise and data is a straight line.

Papers in this version

Showing 60 of 218 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_straight_path_generative_modeling_095056fb,
  title = {Straight Path Generative Modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/straight-path-generative-modeling/rdr_095056fb8eb04291948f754316643fb7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_straight_path_generative_modeling_095056fb_v1,
  title = {Straight Path Generative Modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/straight-path-generative-modeling/rdr_095056fb8eb04291948f754316643fb7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}