meanflow-one-step-generation

name	meanflow-one-step-generation
title	Flow Straighter and Faster: Efficient One-Step Generation via MeanFlow with Rectified Trajectories
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2511.23342
keywords	["flow-matching","one-step-generation","trajectory-learning","diffusion-distillation","generative-models"]
description	Trains mean-velocity models on rectified couplings from pretrained flow models to dramatically smooth loss landscape, enabling faster convergence and superior one-step generation quality without additional training data.

Summary

Flow Straighter and Faster (MeanFlow) addresses one-step generation by identifying that learning mean-velocity fields on highly curved generative trajectories creates noisy optimization. The key innovation trains mean-velocity models on rectified couplings (straighter trajectories from pretrained flow models) combined with distance-based truncation heuristics, dramatically smoothing the loss landscape.

Core Technique

Trajectory Rectification: Instead of learning velocity from independent random couplings (which have arbitrary curvature), use a pretrained flow model to generate "straighter" trajectories. The pretrained model has already learned optimal transports, and sampling from its trajectories provides smoother data for the student model.

Distance-Based Truncation: Remove the most curved trajectory pairs using a distance metric. If two adjacent samples in the coupling show large ||x_end - x_start|| relative to ||z_end - z_start||, they're likely noisy and removed from training.

Mean-Velocity Learning: Train a lightweight model to predict average velocity along trajectories:

v_mean = E[dx/dt] along trajectory

This is simpler than predicting full trajectories and benefits from rectified data.

Implementation

Pretrained flow generation: Load a trained flow matching model (e.g., stable diffusion flow). Sample trajectories for one-step student training.

Trajectory selection: For candidates in the batch, compute:

curvature = ||x_t - x_0|| / (t + ε)  # Distance per unit time
if curvature > threshold:
    skip_sample(trajectory)

Mean-velocity model: Train a small transformer:

velocity = mlp(concat(x_0, x_T, t))  # Predict average velocity
loss = mse(velocity, (x_T - x_0) / (t_max - t_min))

One-step inference: At test time, sample z ~ N(0,I) and apply: x = z + velocity(z, t=0), with optional refinement.

When to Use

• One-step image generation where training data is limited
• Scenarios where distillation from pretrained flow models is available
• Applications requiring both fast inference and good quality
• Tasks where the pretrained model's trajectory space is appropriate

When NOT to Use

• Scenarios requiring multi-step generation for highest quality
• Tasks without access to good pretrained flow models