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

TL;DR

Pixel MeanFlow (pMF) is a one-step image generator that operates directly in raw pixel space — no VAE/latent tokenizer — and is trained from scratch (no distillation). The core recipe decouples the network’s output space (an x-prediction on a presumed low-dimensional image manifold, à la JiT) from the loss space (an average-velocity MeanFlow objective, à la iMF), connected by a closed-form conversion between the three fields x, u (average velocity), and v (instantaneous velocity). On ImageNet 256×256 it reports 2.22 FID and on 512×512 2.48 FID with a single forward pass, closing a gap that had until now only been filled in latent space or with multi-step pixel models.

Key claims

pMF parameterizes a denoised-image-like field x̂ (x-prediction) in pixel space rather than an instantaneous or average velocity, on the assumption that the prediction target lies on a low-dimensional image manifold [§3].
A closed-form conversion relates x, the instantaneous velocity v, and the average velocity u, so a single network can output x̂ while still being supervised by a MeanFlow-style loss in velocity space [§3].
The training objective follows improved MeanFlow (iMF): learn the average velocity field u using a loss defined on the instantaneous-velocity field v [§3].
Decoupling output space from loss space is the central design move — prior MeanFlow / consistency-style methods conflated the two by predicting velocity in the same space they supervised in [§3].
On ImageNet 256×256 class-conditional, 1-NFE FID is 2.22; on 512×512, 1-NFE FID is 2.48 — strong for the latent-free one-step regime [Abstract].
High-resolution training benefits from perceptual losses on top of the MeanFlow objective, and the paper provides ablations on optimizers, time-sampling strategies, and resolution scaling [Abstract / §4].

Method

pMF inherits two prior threads and bridges them. From JiT (Li & He 2025), it takes the “image transformer in raw pixel space” stance — patchify pixels directly, no VAE, and parameterize the network to output x̂ (a denoised image), which is expected to lie on a low-dimensional image manifold. From improved MeanFlow (Geng et al. 2025b), it takes the “average velocity u learned via a loss in instantaneous-velocity v space” objective, which gives one-step sampling without distillation.

The bridge is a conversion: given a noised sample x_t at time t, the fields x (denoised target), u (average velocity from t to 0), and v (instantaneous velocity at t) are related by simple algebra under the chosen noising schedule. The network outputs x̂(x_t, t); the loss converts this to a predicted u via the conversion, and supervises against a v-space MeanFlow target. At inference, one forward pass produces x̂ directly from a noise sample. High-resolution training adds perceptual loss terms.

Results

ImageNet 256×256 class-conditional: 2.22 FID at 1-NFE [Abstract].
ImageNet 512×512 class-conditional: 2.48 FID at 1-NFE [Abstract].
Ablations cover optimizer choice, time-sampling distribution, and the perceptual-loss augmentation needed at 512×512 [§4].

The framing is that pMF “fills a key missing piece” in the one-step-and-latent-free corner of the design space: prior work was either one-step-in-latents (Consistency Models, MeanFlow on SD-VAE) or multi-step-in-pixels (JiT), but not both.

Why it’s interesting

This is the cleanest demonstration to date that one-step generation and pixel-space training are mutually compatible — neither piece has to be sacrificed for the other. The output/loss-space decoupling is a small but pointed idea: prior MeanFlow-style methods supervise velocities directly, which conflates representational capacity (where the network’s output lives) with the training signal (what space the regression target lives in). pMF’s claim is that for pixel-space generation, the manifold assumption belongs on the output side and the MeanFlow consistency belongs on the loss side. For Luma, the implication is that latent-free training stacks remain viable at competitive FID even at 512×512, which removes the VAE from the critical path of any future generator that wants pixel-aware control or pixel-space objectives. The recipe is also a direct sibling of SoFlow: Solution Flow Models for One-Step Generative Modeling in the one-step-from-scratch lineage, but with the orthogonal axis of pixel-space training.