One-step Latent-free Image Generation with Pixel Mean Flows

TL;DR

This work introduces pixel MeanFlow (pMF), a one-step latent-free image generator that outputs a denoised-image-like field $x$ and learns velocity-based objectives through a MeanFlow-based conversion between $x$, $u$, and $v$. By predicting $x$ in pixel space and deriving the corresponding velocity targets, pMF achieves strong Fréchet Inception Distance scores on ImageNet at 256×256 and 512×512 in a single forward pass, demonstrating feasibility of end-to-end pixel-space generation without latent tokens. The authors show that predicting $x$ is crucial at high resolutions, and augment the objective with perceptual losses to further improve quality, while providing thorough ablations on optimizers, time-sampling strategies, and high-resolution scaling. Overall, pMF narrows the gap between one-step and latent-based diffusion/flow methods and highlights the viability of direct end-to-end pixel-space generative modeling.

Abstract

Modern diffusion/flow-based models for image generation typically exhibit two core characteristics: (i) using multi-step sampling, and (ii) operating in a latent space. Recent advances have made encouraging progress on each aspect individually, paving the way toward one-step diffusion/flow without latents. In this work, we take a further step towards this goal and propose "pixel MeanFlow" (pMF). Our core guideline is to formulate the network output space and the loss space separately. The network target is designed to be on a presumed low-dimensional image manifold (i.e., x-prediction), while the loss is defined via MeanFlow in the velocity space. We introduce a simple transformation between the image manifold and the average velocity field. In experiments, pMF achieves strong results for one-step latent-free generation on ImageNet at 256x256 resolution (2.22 FID) and 512x512 resolution (2.48 FID), filling a key missing piece in this regime. We hope that our study will further advance the boundaries of diffusion/flow-based generative models.

Figures (6)

• Figure 1: The pixel MeanFlow (pMF) formulation, driven by the manifold hypothesis.(Left): Following MeanFlow meanflow, pMF aims to approximate the average velocity field$\mathbf{u}(\mathbf{z}_t, r, t)$ induced by the underlying ODE trajectory (black). We define a new field $\mathbf{x}(\mathbf{z}_t, r, t) \triangleq \mathbf{z}_t - t\cdot\mathbf{u}(\mathbf{z}_t, r, t)$, which behaves like denoised images. We hypothesize that $\mathbf{x}$ approximately lies on a low-dimensional data manifold (orange curve) and can therefore be more accurately approximated by a neural network. (Right): Visualization of the quantities $\mathbf{z}_t$, $\mathbf{u}$, $\mathbf{x}$ obtained by tracking an ODE trajectory via simulation. The average velocity field $\mathbf{u}$ corresponds to noisy images and is inevitably higher-dimensional; the induced field $\mathbf{x}$ corresponds to approximately clean or blurred images, which can be easier to model by a neural network.
• Figure 2: Toy Experiment. A 2D toy dataset is linearly projected into a $D$-dimensional observation space using a fixed, $D{\times}2$ column-orthonormal projection matrix. We train MeanFlow models with either the original $\mathbf{u}$-prediction or the proposed $\mathbf{x}$-prediction, for $D \in \{2, 8, 16, 512\}$. We visualize 1-NFE generation results. The models use the same 7-layer ReLU MLP backbone with 256 hidden units. The $\mathbf{x}$-prediction formulation produces reasonably good results, whereas $\mathbf{u}$-prediction fails in the case of high-dimensional observation spaces.
• Figure 3: Training curves of pMF on ImageNet 256$\times$256 with pixel-space, 1-NFE generation.
• Figure 4: Qualitative results of 1-NFE pixel-space generation on ImageNet 256$\times$256. We show uncurated results of pMF-H/16 on the five classes listed here; more are in Appendix \ref{['app:vis']}.
• Figure 5: Uncurated 1-NFE pixel class-conditional generation samples of pMF-H/16 on ImageNet 256$\times$256.